Updated October 31, 2026
Upstream bugs
Tool calls also break in the inference servers. Each model's tool-calling bugs in vLLM, SGLang, TensorRT-LLM and XGrammar: what's open, and how long fixes take to ship.
12open tool-calling issues upstreamvLLM 6 · SGLang 4 · TensorRT-LLM 2 · XGrammar 0 · oldest open 42 days
- 0112 tool-calling bugs are open upstream for DeepSeek V4.1 Flash 1017 (vLLM 6, SGLang 4, TensorRT-LLM 2), the oldest for 42 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now12oldest 42 days
Opened / fixed12 / 0fixed: in a release
First server supportday −32XGrammar; vLLM day −25; SGLang day −15; TensorRT-LLM day 5
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (58)
- day 13TensorRT-LLM #21072fix[None][fix] Responses: detokenize for the parsers as chat completions doesmerged, not released
- day 13TensorRT-LLM #21104fix[None][fix] DeepSeek-V4.1: namespaced Responses tools and calls in the checkpoint's formmerged, not released
- day 12SGLang #47167bug[Bug] Tools + JSON-schema format cannot be combined: Kimi-K3 returns 400, glm47 / DeepSeek silently drop tool calls and invent the answer (breaks Codex --output-schema)open
- day 12vLLM #64433bug[Bug]: vllm-rs render drops tools for tool_choice "none" and rejects empty content, unlike the Python frontendopen
- day 12vLLM #58780fix[Bugfix][Parser] Stream DeepSeek V4/V3.2 string tool arguments incrementallymerged, not released
- day 12SGLang #46780fix[Feat] rust: Parse DeepSeek-V4/V4.1 output with deepseek-recipemerged, not released
- day 10TensorRT-LLM #20873fix[None][fix] DeepSeek DSML tool parsers: the layout newlines around tool calls are not contentin v1.4.0rc2
- day 10vLLM #63762fix[Rust Frontend] Separate DeepSeek V4 content blocks like deepseek-recipemerged, not released
- day 10TensorRT-LLM #21025fix[None][fix] DeepSeek DSML tool parsers: keep string parameter values as writtenin v1.4.0rc2
- day 10TensorRT-LLM #21051fix[None][fix] DeepSeek renderers: tools join the leading system message; V4.1 renders developer as systemin v1.4.0rc2
- day 10TensorRT-LLM #21056fix[None][fix] Responses: resolve namespace::tool calls and restore a dropped namespace on replayin v1.4.0rc2
- day 10SGLang #46706fix[Fix] dsv4: Render null contents, trailing tool-call turns and numeric string efforts with deepseek-recipemerged, not released
- day 9TensorRT-LLM #20989bug[Performance]: DeepSeek streaming tool parsers rescan the whole tool call on every increment (30 s of CPU for one 64k-character call)open
- day 9TensorRT-LLM #21014bug[Bug]: DeepSeek V4.1 / V3.2 tool parsers cut a parameter value at a literal DSML closing tag and drop or leak the rest of the callopen
- day 9vLLM #63794bug[Bug]: Persian half-spaces (U+200C) are removed from DeepSeek-V4.1-Flash-1017 responses when using toolsopen
- day 8SGLang #46348bug[Bug] dsv41: deepseek-recipe rendering differs from the checkpoint's encoding.py (tool namespaces, current-turn reasoning before tool results, action)open
- day 7SGLang #45315feature[Feat] openai: Render DeepSeek-V4/V4.1 chat prompts with deepseek-recipemerged, not released
- day 7SGLang #45907support[Feat] openai: parse DeepSeek output through deepseek-recipemerged, not released
- day 7SGLang #46219feature[Feat] openai: Parse DeepSeek text for the native endpoints through deepseek-recipemerged, not released
- day 6vLLM #63314fix[Rust Frontend] Render empty tool call arguments as an empty object like Pythonmerged, not released
- day 5TensorRT-LLM v1.4.0rc1supportmodel supportin v1.4.0rc1
- day 5vLLM #62875fix[Rust Frontend] Match DeepSeek DSML markers by token identitymerged, not released
- day 4vLLM #62741feature[Rust Frontend] Support --exclude-tools-when-tool-choice-nonemerged, not released
- day 2vLLM #60792fix[Bugfix][Frontend] Align DeepSeek V4.1 images with rendered contentmerged, not released
- day 2vLLM #62199fix[Bugfix][Frontend] Match DeepSeek-V4.1 encoder to the 1017 reference encodingmerged, not released
- day 2vLLM #62200fix[Bugfix][Rust Frontend] Match DeepSeek-V4.1 renderer to the 1017 reference encodingmerged, not released
- day 1SGLang #44941bugDeepSeek V3.2/V4/V4.1 streaming tool-call parser: pre-existing bugs found while fuzzingopen
- day −1vLLM #62051bug[Bug]: DeepSeek V4.1 tool parser drops malformed DSML blocks and emits invalid/empty args on truncationopen
- day −1SGLang #44751fix[Fix] dsv41: Preserve tool-result image order and current-turn thinkingin v0.5.22
- day −3SGLang #43625support[sgl-router] Render GLM-5.3 and GLM-5.3-Flash through sglang-processorin v0.5.22
- day −3SGLang #44243fix[Fix] sglang-processor: Follow Python's OpenAI layer through #43027, #37577, #43006, #43026, #43005in v0.5.22
- day −4SGLang #43006fixfix(openai): honor tool_choice=none in DeepSeek encodersin v0.5.22
- day −5vLLM #60781fix[Rust Frontend] Compose answer constraints into parser-owned output grammarsin v0.32.0
- day −6TensorRT-LLM #19727support[TRTLLM-16621][feat] Support DeepSeek-V4.1-Flashin v1.4.0rc0
- day −7SGLang #43000support[sgl-router] Serve DeepSeek-V4.1 Flash through sglang-processorin v0.5.22
- day −9vLLM #59627fix[Bugfix][Structured Output] Feed the token that implicitly ends reasoning to the grammarin v0.32.0
- day −15SGLang v0.5.21supporttool-calling supportin v0.5.21
- day −17vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day −19SGLang #41226fix[sgl-router] Scope input_ids forwarding by renderer: all text chats for DeepSeek-V4in v0.5.21
- day −20SGLang #41449bug[Bug] DSpark + TP: grammar-constrained request batched with any other request deadlocks all ranks (overlap and non-overlap), GPUs spin in a collectiveopen
- day −20XGrammar #936fixfix: bound DeepSeek V4.1 auto tool callsmerged, not released
- day −22vLLM #58640bug[Bug]: When using the V4 parser in vLLM to parse outputs from DeepSeek V4.1 Flash, the leading extra space in V4.1 DSML tags breaks state machine matching and parameter regex, filters arg deltas, and prevents streaming return of tool call arguments.open
- day −25vLLM v0.30.0supportmodel supportin v0.30.0
- day −26SGLang #39026fixfeat: use XGrammar V4.1 DSML parameter constraintsin v0.5.21
- day −27vLLM #57826bug[Bug]: minimax_m2 / glm47_moe / deepseek_v4 arg converters drop the last parameter when the model omits the closing tag (same class as #57699)open
- day −28vLLM #57730bug[Bug]: /tokenize returns 400 "cannot pickle ValidatorIterator" for assistant `content: null` + `tool_calls` (DeepSeek-V4.1), while chat completions accepts the same messagesopen
- day −29vLLM #57272fix[Frontend] Upgrade XGrammar to 0.2.7 and Rust structural tags to 0.3.0in v0.30.0
- day −30SGLang #39665supportdsv4.1: chat encoding and tool parsingin v0.5.20
- day −30XGrammar #900fixBugfix: improve DeepSeek V4.1 tool-call performance and grammar correctnessin v0.2.8
- day −31XGrammar #884otherrefactor: split JSON schema converter extensions by modelin v0.2.8
- day −32XGrammar v0.2.7supporttool-calling supportin v0.2.7
- day −32vLLM #56408fix[Frontend] Use XGrammar schema constraints for DeepSeek V4.1in v0.30.0
- day −33XGrammar #885supportfeat: support DeepSeek V4.1 structural tag formatin v0.2.7
- day −33XGrammar #886fixfix: correlate DeepSeek V4.1 parameter attributes with value typesin v0.2.7
- day −33XGrammar #889fixfix: correlate DeepSeek V4.1 parameter attributes with value typesin v0.2.7
- day −36vLLM #56260fix[Bugfix][Rust Frontend][Renderer] Align DeepSeek tool-call arguments with deepseek-recipein v0.30.0
- day −37vLLM #56208support[Model][Frontend] Support DeepSeek-V4.1-Flash in Rust and Python frontendsin v0.30.0
- day −37SGLang #38951fixDeepSeek V4.1: build the tool-call structural tag with the spaced DSML namesin v0.5.20
11open tool-calling issues upstreamvLLM 6 · SGLang 5 · TensorRT-LLM 0 · XGrammar 0 · oldest open 55 days
- 0111 tool-calling bugs are open upstream for GLM 5.3 Flash (vLLM 6, SGLang 5), the oldest for 55 days.
- 02vLLM v0.32.0 fixed #59608 ([Bug]: Structured output skips the `<tool_call>` that implicitly ends reasoning, so GLM-4.7/5.x `required`/named tool calls are lost) 13 days ago: providers on an older version still have the bug.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix11 daysreport to release
Open now11oldest 55 days
Opened / fixed14 / 3fixed: in a release
First server supportday 23SGLang; vLLM day 27; TensorRT-LLM day 49
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (31)
- day 65vLLM #62424support[Rust Frontend] Add native GLM-5.3 / GLM-5.3-Flash chat renderingmerged, not released
- day 65TensorRT-LLM #21109fix[None][fix] GLM reasoning parser: drop the owed </think> after a tool call; read the thinking mode off the promptmerged, not released
- day 65vLLM #64615fix[Bugfix][Frontend] Leave GLM-4.7 array and object tool arguments unpinnedmerged, not released
- day 64SGLang #47167bug[Bug] Tools + JSON-schema format cannot be combined: Kimi-K3 returns 400, glm47 / DeepSeek silently drop tool calls and invent the answer (breaks Codex --output-schema)open
- day 64vLLM #64535fix[Bugfix][Parser] Parse closed GLM-4.7 tool arguments once while streamingmerged, not released
- day 63vLLM #64326bug[Bug]: DelegatingParser.parse ignores model_output_token_ids, so markers written as plain text become tool calls or end reasoning (non-streaming)open
- day 61SGLang #45659fix[Fix] rust-server: Match Python's chat layer on templates, streaming and lowered-chat errorsmerged, not released
- day 58SGLang #45953bug[Bug] Glm47MoeDetector.detect_and_parse drops tool calls when the function name is followed by a newline (non-streaming)open
- day 57vLLM #63185bug[Bug]: Responses API rejects `include: ["web_search_call.results", "web_search_call.action.sources"]` with 400, although the `web_search` tool is acceptedopen
- day 55SGLang #41233fix[Fix] glm47: Disable tool grammar for `tool_choice="none"`merged, not released
- day 55SGLang #45154feature[Feat] glm-5.3: Render GLM-5.3 and GLM-5.3-Flash chat prompts with a Python chat templatemerged, not released
- day 54SGLang #45159bug[Bug] GLM-5.3-Flash on 8×H20 (TP8/EP8) with EAGLE: CUDA illegal memory access in run_eagle_verify → build_grammar_vocab_maskopen
- day 54vLLM #62483bug[Bug][Structured Outputs] Parser._apply_structural_tag overrides a request's response_format when tools are present — schema not enforced (GLM-5.3)open
- day 51vLLM #62051bug[Bug]: DeepSeek V4.1 tool parser drops malformed DSML blocks and emits invalid/empty args on truncationopen
- day 51vLLM #61574fix[Bugfix][Structured Output] Mask grammar-invalid implicit reasoning endsin v0.32.0
- day 49TensorRT-LLM v1.4.0rc0supportmodel supportin v1.4.0rc0
- day 49SGLang #43625support[sgl-router] Render GLM-5.3 and GLM-5.3-Flash through sglang-processorin v0.5.22
- day 47vLLM #57322feature[Frontend] Support CustomToolCall in the non-Harmony Responses APIin v0.32.0
- day 43SGLang #43085bug[Bug]open
- day 43vLLM #59627fix[Bugfix][Structured Output] Feed the token that implicitly ends reasoning to the grammarin v0.32.0
- day 37SGLang #42269bugresponse_format + tools on glm47: tool calls silently dropped, model invents a JSON answeropen
- day 36vLLM #59608bug[Bug]: Structured output skips the `<tool_call>` that implicitly ends reasoning, so GLM-4.7/5.x `required`/named tool calls are lostfixed in v0.32.0
- day 36SGLang #39273fix[AMD] [GLM-5.3-Flash] Enable FP8 and MXFP4 serving on gfx950in v0.5.21
- day 27vLLM v0.30.0supportmodel supportin v0.30.0
- day 27SGLang #40497fix[Docs] GLM-5.3/5.3-Flash cookbooks: enable reasoning/tool-call parsers by default via autoin v0.5.21
- day 23SGLang v0.5.20supportmodel supportin v0.5.20
- day 21SGLang #39782bug[Bug] Anthropic streaming merges multiple tool-call arguments into one tool_use blockfixed in v0.5.20
- day 20vLLM #57087bug[Bug] GLM-5.3 hybrid (nvfp4, GB10/sm121, TP2): concurrent batching corrupts non-ASCII generation — fragment-token state poisoning + silent DecodeStream FFFD flushopen
- day 16SGLang #38297fixAuto-detect GLM-5.3 chat templates as glm45/glm47 parsersin v0.5.20
- day 11vLLM #55541bug[Bug]: GLM-5.3-Flash with forced tool_choice (named function) fails to converge — runs to max_tokens and returns ▎ truncated tool_call.argumentsopen
- day 3vLLM #54337bug[Bug]: Assistant `content=null` + `tool_calls` renders literal "None" into chat template context — degenerate outputs in agent workloadsfixed in v0.29.0
12open tool-calling issues upstreamvLLM 6 · SGLang 4 · TensorRT-LLM 2 · XGrammar 0 · oldest open 42 days
- 0112 tool-calling bugs are open upstream for DeepSeek V4.1 Flash (vLLM 6, SGLang 4, TensorRT-LLM 2), the oldest for 42 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now12oldest 42 days
Opened / fixed12 / 0fixed: in a release
First server supportday 5XGrammar; vLLM day 12; SGLang day 22; TensorRT-LLM day 42
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (58)
- day 50TensorRT-LLM #21072fix[None][fix] Responses: detokenize for the parsers as chat completions doesmerged, not released
- day 50TensorRT-LLM #21104fix[None][fix] DeepSeek-V4.1: namespaced Responses tools and calls in the checkpoint's formmerged, not released
- day 49SGLang #47167bug[Bug] Tools + JSON-schema format cannot be combined: Kimi-K3 returns 400, glm47 / DeepSeek silently drop tool calls and invent the answer (breaks Codex --output-schema)open
- day 49vLLM #64433bug[Bug]: vllm-rs render drops tools for tool_choice "none" and rejects empty content, unlike the Python frontendopen
- day 49vLLM #58780fix[Bugfix][Parser] Stream DeepSeek V4/V3.2 string tool arguments incrementallymerged, not released
- day 49SGLang #46780fix[Feat] rust: Parse DeepSeek-V4/V4.1 output with deepseek-recipemerged, not released
- day 47TensorRT-LLM #20873fix[None][fix] DeepSeek DSML tool parsers: the layout newlines around tool calls are not contentin v1.4.0rc2
- day 47vLLM #63762fix[Rust Frontend] Separate DeepSeek V4 content blocks like deepseek-recipemerged, not released
- day 47TensorRT-LLM #21025fix[None][fix] DeepSeek DSML tool parsers: keep string parameter values as writtenin v1.4.0rc2
- day 47TensorRT-LLM #21051fix[None][fix] DeepSeek renderers: tools join the leading system message; V4.1 renders developer as systemin v1.4.0rc2
- day 47TensorRT-LLM #21056fix[None][fix] Responses: resolve namespace::tool calls and restore a dropped namespace on replayin v1.4.0rc2
- day 47SGLang #46706fix[Fix] dsv4: Render null contents, trailing tool-call turns and numeric string efforts with deepseek-recipemerged, not released
- day 46TensorRT-LLM #20989bug[Performance]: DeepSeek streaming tool parsers rescan the whole tool call on every increment (30 s of CPU for one 64k-character call)open
- day 46TensorRT-LLM #21014bug[Bug]: DeepSeek V4.1 / V3.2 tool parsers cut a parameter value at a literal DSML closing tag and drop or leak the rest of the callopen
- day 46vLLM #63794bug[Bug]: Persian half-spaces (U+200C) are removed from DeepSeek-V4.1-Flash-1017 responses when using toolsopen
- day 45SGLang #46348bug[Bug] dsv41: deepseek-recipe rendering differs from the checkpoint's encoding.py (tool namespaces, current-turn reasoning before tool results, action)open
- day 44SGLang #45315feature[Feat] openai: Render DeepSeek-V4/V4.1 chat prompts with deepseek-recipemerged, not released
- day 44SGLang #45907support[Feat] openai: parse DeepSeek output through deepseek-recipemerged, not released
- day 44SGLang #46219feature[Feat] openai: Parse DeepSeek text for the native endpoints through deepseek-recipemerged, not released
- day 43vLLM #63314fix[Rust Frontend] Render empty tool call arguments as an empty object like Pythonmerged, not released
- day 42TensorRT-LLM v1.4.0rc1supportmodel supportin v1.4.0rc1
- day 42vLLM #62875fix[Rust Frontend] Match DeepSeek DSML markers by token identitymerged, not released
- day 41vLLM #62741feature[Rust Frontend] Support --exclude-tools-when-tool-choice-nonemerged, not released
- day 39vLLM #60792fix[Bugfix][Frontend] Align DeepSeek V4.1 images with rendered contentmerged, not released
- day 39vLLM #62199fix[Bugfix][Frontend] Match DeepSeek-V4.1 encoder to the 1017 reference encodingmerged, not released
- day 39vLLM #62200fix[Bugfix][Rust Frontend] Match DeepSeek-V4.1 renderer to the 1017 reference encodingmerged, not released
- day 38SGLang #44941bugDeepSeek V3.2/V4/V4.1 streaming tool-call parser: pre-existing bugs found while fuzzingopen
- day 36vLLM #62051bug[Bug]: DeepSeek V4.1 tool parser drops malformed DSML blocks and emits invalid/empty args on truncationopen
- day 36SGLang #44751fix[Fix] dsv41: Preserve tool-result image order and current-turn thinkingin v0.5.22
- day 34SGLang #43625support[sgl-router] Render GLM-5.3 and GLM-5.3-Flash through sglang-processorin v0.5.22
- day 34SGLang #44243fix[Fix] sglang-processor: Follow Python's OpenAI layer through #43027, #37577, #43006, #43026, #43005in v0.5.22
- day 33SGLang #43006fixfix(openai): honor tool_choice=none in DeepSeek encodersin v0.5.22
- day 32vLLM #60781fix[Rust Frontend] Compose answer constraints into parser-owned output grammarsin v0.32.0
- day 31TensorRT-LLM #19727support[TRTLLM-16621][feat] Support DeepSeek-V4.1-Flashin v1.4.0rc0
- day 30SGLang #43000support[sgl-router] Serve DeepSeek-V4.1 Flash through sglang-processorin v0.5.22
- day 28vLLM #59627fix[Bugfix][Structured Output] Feed the token that implicitly ends reasoning to the grammarin v0.32.0
- day 22SGLang v0.5.21supporttool-calling supportin v0.5.21
- day 20vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 18SGLang #41226fix[sgl-router] Scope input_ids forwarding by renderer: all text chats for DeepSeek-V4in v0.5.21
- day 17SGLang #41449bug[Bug] DSpark + TP: grammar-constrained request batched with any other request deadlocks all ranks (overlap and non-overlap), GPUs spin in a collectiveopen
- day 17XGrammar #936fixfix: bound DeepSeek V4.1 auto tool callsmerged, not released
- day 15vLLM #58640bug[Bug]: When using the V4 parser in vLLM to parse outputs from DeepSeek V4.1 Flash, the leading extra space in V4.1 DSML tags breaks state machine matching and parameter regex, filters arg deltas, and prevents streaming return of tool call arguments.open
- day 12vLLM v0.30.0supportmodel supportin v0.30.0
- day 11SGLang #39026fixfeat: use XGrammar V4.1 DSML parameter constraintsin v0.5.21
- day 10vLLM #57826bug[Bug]: minimax_m2 / glm47_moe / deepseek_v4 arg converters drop the last parameter when the model omits the closing tag (same class as #57699)open
- day 9vLLM #57730bug[Bug]: /tokenize returns 400 "cannot pickle ValidatorIterator" for assistant `content: null` + `tool_calls` (DeepSeek-V4.1), while chat completions accepts the same messagesopen
- day 8vLLM #57272fix[Frontend] Upgrade XGrammar to 0.2.7 and Rust structural tags to 0.3.0in v0.30.0
- day 7SGLang #39665supportdsv4.1: chat encoding and tool parsingin v0.5.20
- day 7XGrammar #900fixBugfix: improve DeepSeek V4.1 tool-call performance and grammar correctnessin v0.2.8
- day 6XGrammar #884otherrefactor: split JSON schema converter extensions by modelin v0.2.8
- day 5XGrammar v0.2.7supporttool-calling supportin v0.2.7
- day 5vLLM #56408fix[Frontend] Use XGrammar schema constraints for DeepSeek V4.1in v0.30.0
- day 4XGrammar #885supportfeat: support DeepSeek V4.1 structural tag formatin v0.2.7
- day 4XGrammar #886fixfix: correlate DeepSeek V4.1 parameter attributes with value typesin v0.2.7
- day 4XGrammar #889fixfix: correlate DeepSeek V4.1 parameter attributes with value typesin v0.2.7
- day 1vLLM #56260fix[Bugfix][Rust Frontend][Renderer] Align DeepSeek tool-call arguments with deepseek-recipein v0.30.0
- day 0vLLM #56208support[Model][Frontend] Support DeepSeek-V4.1-Flash in Rust and Python frontendsin v0.30.0
- day 0SGLang #38951fixDeepSeek V4.1: build the tool-call structural tag with the spaced DSML namesin v0.5.20
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now0
Opened / fixed0 / 0fixed: in a release
First server supportday 12vLLM; SGLang day 21
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (5)
- day 25vLLM #57340fix[Rust Frontend] Build full-output grammars from initialized reasoning parsersin v0.31.0
- day 21SGLang v0.5.20supportmodel supportin v0.5.20
- day 12vLLM v0.29.0supportmodel supportin v0.29.0
- day 8SGLang #36805supportSupport Hy4-preview (#36805)in v0.5.19
- day 1vLLM #54160support[Hy4] support Hy4-preview modelin v0.29.0
19open tool-calling issues upstreamvLLM 10 · SGLang 7 · TensorRT-LLM 2 · XGrammar 0 · oldest open 115 days
- 0119 tool-calling bugs are open upstream for DeepSeek V4 Flash 0731 (vLLM 10, SGLang 7, TensorRT-LLM 2), the oldest for 115 days.
- 02SGLang v0.5.22 fixed #33163 ([Bug] deepseek-v4-flash toolcall error runner_backend from marlin to flashinfer_mxfp4) 12 days ago: providers on an older version still have the bug.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix12 daysreport to release
Open now19oldest 115 days
Opened / fixed40 / 21fixed: in a release
First server supportday −95vLLM; XGrammar day −93; SGLang day −76; TensorRT-LLM day −71
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (105)
- day 91TensorRT-LLM #21072fix[None][fix] Responses: detokenize for the parsers as chat completions doesmerged, not released
- day 90vLLM #58780fix[Bugfix][Parser] Stream DeepSeek V4/V3.2 string tool arguments incrementallymerged, not released
- day 90SGLang #46780fix[Feat] rust: Parse DeepSeek-V4/V4.1 output with deepseek-recipemerged, not released
- day 89vLLM #64326bug[Bug]: DelegatingParser.parse ignores model_output_token_ids, so markers written as plain text become tool calls or end reasoning (non-streaming)open
- day 88TensorRT-LLM #20873fix[None][fix] DeepSeek DSML tool parsers: the layout newlines around tool calls are not contentin v1.4.0rc2
- day 88vLLM #63762fix[Rust Frontend] Separate DeepSeek V4 content blocks like deepseek-recipemerged, not released
- day 88TensorRT-LLM #21025fix[None][fix] DeepSeek DSML tool parsers: keep string parameter values as writtenin v1.4.0rc2
- day 88TensorRT-LLM #21051fix[None][fix] DeepSeek renderers: tools join the leading system message; V4.1 renders developer as systemin v1.4.0rc2
- day 88SGLang #46706fix[Fix] dsv4: Render null contents, trailing tool-call turns and numeric string efforts with deepseek-recipemerged, not released
- day 87TensorRT-LLM #20989bug[Performance]: DeepSeek streaming tool parsers rescan the whole tool call on every increment (30 s of CPU for one 64k-character call)open
- day 86SGLang #46348bug[Bug] dsv41: deepseek-recipe rendering differs from the checkpoint's encoding.py (tool namespaces, current-turn reasoning before tool results, action)open
- day 85SGLang #45315feature[Feat] openai: Render DeepSeek-V4/V4.1 chat prompts with deepseek-recipemerged, not released
- day 85SGLang #45907support[Feat] openai: parse DeepSeek output through deepseek-recipemerged, not released
- day 85SGLang #46219feature[Feat] openai: Parse DeepSeek text for the native endpoints through deepseek-recipemerged, not released
- day 84vLLM #63314fix[Rust Frontend] Render empty tool call arguments as an empty object like Pythonmerged, not released
- day 83vLLM #62875fix[Rust Frontend] Match DeepSeek DSML markers by token identitymerged, not released
- day 82vLLM #62741feature[Rust Frontend] Support --exclude-tools-when-tool-choice-nonemerged, not released
- day 79SGLang #44941bugDeepSeek V3.2/V4/V4.1 streaming tool-call parser: pre-existing bugs found while fuzzingopen
- day 75SGLang #43625support[sgl-router] Render GLM-5.3 and GLM-5.3-Flash through sglang-processorin v0.5.22
- day 75SGLang #44243fix[Fix] sglang-processor: Follow Python's OpenAI layer through #43027, #37577, #43006, #43026, #43005in v0.5.22
- day 70SGLang #42999support[sgl-router] Parse tool calls when serving chat through /generatein v0.5.22
- day 69vLLM #60722bug[RFC]: Strcutured Outputs Not Respected When Reasoning Implicitly Endsfixed day 81, not released
- day 69vLLM #59627fix[Bugfix][Structured Output] Feed the token that implicitly ends reasoning to the grammarin v0.32.0
- day 69SGLang #42664fix[rust-processor] DeepSeek-V4 parity with SGLang, plus a reusable parity harness and skillin v0.5.22
- day 65vLLM #59926bug[Bug] DeepSeek-V4-Flash tool-call format failures associated with prefix-cache reuse; worst for short cached head + long uncached suffix; `cache_salt` reduces failures but is confounded with hit shape (v0.27.1, fp8 KV, MTP)open
- day 61vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 59SGLang #41226fix[sgl-router] Scope input_ids forwarding by renderer: all text chats for DeepSeek-V4in v0.5.21
- day 58XGrammar #936fixfix: bound DeepSeek V4.1 auto tool callsmerged, not released
- day 57SGLang #41317bug[Bug] DeepSeek V3.2/V4 DSML detector strips whitespace (e.g. trailing newline) from string="true" parameter valuesopen
- day 54vLLM #58280bug[Bug]: DeepSeek-V4 chat template puts `tools` into a synthetic leading system message instead of on the existing system message (diverges from the reference encoder)fixed in v0.31.0
- day 53SGLang #39632fixfix(function_call): buffer complete DeepSeek DSML invokesin v0.5.21
- day 51vLLM #57826bug[Bug]: minimax_m2 / glm47_moe / deepseek_v4 arg converters drop the last parameter when the model omits the closing tag (same class as #57699)open
- day 49SGLang #40236bug[Bug] DeepSeek V4/V3.2: DSML tool calls returned as content with no tool_calls (bare invoke, unterminated section, malformed sibling)open
- day 44SGLang #39277bug[Bug] b12x-vision dev-v4f-2dgx regresses DSV4 history argument encoding from #28035open
- day 42vLLM #56482bug[Bug]: DSML tool-call recovery (#55954) emits ghost empty-argument calls on truncation and narration, and recovers undeclared toolsopen
- day 42vLLM #56260fix[Bugfix][Rust Frontend][Renderer] Align DeepSeek tool-call arguments with deepseek-recipein v0.30.0
- day 41vLLM #56141fix[Bugfix] Tolerate misspelled DSML tool_calls wrapperin v0.30.0
- day 40vLLM #55954fix[Bugfix] Parse DSML tool calls when the model omits the tool_calls wrapperin v0.29.0
- day 39vLLM #55223fix[Perf] Eliminate full-history reasoning scans for structured outputsin v0.29.0
- day 33vLLM #54838fix[Bugfix] Implicitly close DeepSeek DSML parametersin v0.29.0
- day 32vLLM #53281fix[Bugfix][Rust Frontend] Fix adjacent DeepSeek V4 user content renderingin v0.29.0
- day 26vLLM #53831bug[Bug]: DeepSeek-V4-Flash tool_choice=auto silently drops tool calls at deep reasoning_effort (xhigh/max) — distinct from #41132/#51914fixed in v0.28.0
- day 26vLLM #48922fix[Bugfix] Guard tool call argument JSON parsing in chat message postprocessingin v0.29.0
- day 21vLLM #53046fix[Bugfix][Structured Output] Avoid spurious FSM errors after speculative reasoning endin v0.28.0
- day 20SGLang #35736bug[Bug] DeepSeek-V4 leaks raw DSML when tool_choice="none"open
- day 19vLLM #52846bug[Bug]: DeepSeek V4 buffers long string tool arguments until </parameter>fixed day 90, not released
- day 19SGLang #35563bug[Bug] deepseekv32/deepseekv4 tool-call parsers drop calls when the streamed output has two chunksfixed in v0.5.18
- day 19TensorRT-LLM #17903fix[#17580][fix] Stream the text that precedes a tool call in DeepSeek parsersin v1.3.0rc25
- day 18TensorRT-LLM #17573fix[#17572][fix] Emit the withheld buffer in DeepSeek streaming tool parsersin v1.3.0rc25
- day 12vLLM #51914bug[Bug] DeepSeek-V4-Flash-0731 intermittently emits malformed DSML tool-call start wrapper on v0.27.1 + DSparkopen
- day 12TensorRT-LLM #17572bug[Bug]: DeepSeek streaming tool parsers drop text they withheld while waiting on a tool-call tokenfixed in v1.3.0rc25
- day 12TensorRT-LLM #17580bug[Bug]: DeepSeek streaming tool parsers drop text that precedes a tool call in the same deltafixed in v1.3.0rc25
- day 12SGLang #34458fix[Fix] Make DeepSeek-V4 reasoning and tool-call streaming parsing chunk-invariantin v0.5.18
- day 11vLLM #51829bug[Bug]: DeepSeek V4 Python renderer misplaces tools when a system message existsfixed in v0.30.0
- day 11vLLM #51144feature[Rust Frontend] Support dynamic tools from developer messagesin v0.27.1
- day 10SGLang #34214bug[Bug] Streaming response truncates content when both content and tool_calls are present (SGLang 0.5.17)fixed day 85, not released
- day 10vLLM #51296fix[Bugfix] Align deepseek v4 parser thinking default with tokenizerin v0.27.0
- day 10vLLM #51727fix[Bugfix] Fix DeepSeek V4/3.2 tokenizer vocab size overcount crashing guided decodingin v0.27.1
- day 8vLLM #51467bug[Bug]: DeepSeek-V4-Flash-0731 `response_format` (structured output) crashes the vLLM EngineCore — `apply_grammar_bitmask` tensor size mismatch (4040 vs 4041)fixed in v0.27.1
- day 6SGLang #33867bug[Bug] /v1/responses API fails with 400 when function_call_output output is an arrayfixed in v0.5.22
- day 3vLLM #50924bug[Bug]: EngineCore dies on first guided-decoding request when dspark speculative decoding is enabled (grammar bitmask width mismatch)fixed in v0.27.1
- day 1SGLang #33163bug[Bug] deepseek-v4-flash toolcall error runner_backend from marlin to flashinfer_mxfp4fixed in v0.5.22
- day −2TensorRT-LLM #17021bugDeepSeek-V4: OpenAI `tools` field reorders the system prompt after the tool schemasopen
- day −6SGLang #31753fix[DSPARK] Grammar-constrained decoding, incl. tool_choice=autoin v0.5.17
- day −13vLLM #49002bug[Bug]: Speculative Decoding + Structured Output（tool call）组合下，decode 阶段出现秒级卡顿fixed in v0.30.0
- day −14vLLM #48931bug[Bug] DeepSeek-V4 tool-call parser leaks raw DSML into content when the model omits the `<｜DSML｜tool_calls>` START token (long context)fixed in v0.29.0
- day −21vLLM #48207bug[Bug]: Analysis for an Automated Request Validation and Normalization Module at the Upstream of vLLM (Hard Interception) to Fundamentally Solve DSML Leakage, Garbled Output, and Empty Responsesopen
- day −21SGLang #29883fix[BUG] fix strip streaming empty-string suffix from DSV4 tool argumentsin v0.5.15
- day −22vLLM #48089bug[Bug]: DeepSeek V4 reasoning/tool parsers produce unbalanced tags under concurrent loadfixed in v0.29.0
- day −22SGLang #30623other[Refactor] Share chat encoding dispatch between serving and offline toolsin v0.5.15
- day −23vLLM #47974bug[Bug]: vllm 校验arguments合法性open
- day −23vLLM #47986bug[Bug]: DeepSeek V4/V32 parser can unwrap arguments using the wrong tool's schema when parallel tool calls share identical raw argument textopen
- day −27vLLM #45877fix[Frontend] [Parser] Port DeepSeek V4 to streaming parser engine frameworkin v0.25.0
- day −33TensorRT-LLM #15414support[None][feat] DSv4: model, tokenizer, and integration coveragein v1.3.0rc20
- day −37XGrammar #671bug[Bug] deepseek_xml structural tag: regex pattern starting with an empty alternative (e.g. ^$|...) produces invalid EBNF (Expect element, but got |)fixed in v0.2.4
- day −38SGLang #28449supportExpand parser auto detection coveragein v0.5.14
- day −63SGLang #25676supportUpgrade xgrammar to 0.2.1in v0.5.13
- day −64vLLM #42879fix[Bugfix] Stream DeepSeek DSML tool-call argument deltas incrementallyin v0.22.0
- day −65vLLM #43582fix[Rust Frontend] Add reasoning/tool parser & renderer roundtrip testsin v0.22.0
- day −65vLLM #43662fix[Rust Frontend] Align tool parser fallback behavior between streaming & non-streaming pathsin v0.22.0
- day −69SGLang #26164fix[docs] DeepSeek-V4 cookbook: balanced MegaMoE cap, H200 Pro FP4 mem-frac, nsa-* compat, PD-disagg fixesin v0.5.12.post1
- day −71TensorRT-LLM v1.3.0rc15supportmodel supportin v1.3.0rc15
- day −71vLLM #43255other[CI] Add composed-schema regression tests for DeepSeek V3.2/V4 parsersin v0.22.0
- day −72SGLang #24251fix[RL][TITO] Preserve whitespace in reasoning parser outputsin v0.5.12.post1
- day −75vLLM #42878bug[Bug]: DeepSeek V4 DSML tool calls fake-stream arguments instead of incremental deltasfixed in v0.22.0
- day −75XGrammar #638fixfix(structural_tag): align deepseek_v3_2/v4 invoke separator with chat templatein v0.2.1
- day −76SGLang v0.5.12supporttool-calling supportin v0.5.12
- day −78XGrammar #637bugdeepseek_v3_2 structural tag: double-newline between invoke blocks blocks parallel tool calls at greedy decodingfixed in v0.2.1
- day −83SGLang #24793support[DSV4] Cherry pick missing commits from deepseek_v4 branch and enhance tests (#24793)in v0.5.12
- day −84SGLang #23882supportDeepseek V4 (#23882)in v0.5.12
- day −86vLLM #41801fix[Bugfix] DeepSeekV32/v4: respect string='true|false' attribute andunwrap arguments/input wrapperin v0.20.2
- day −91XGrammar #610fixfix: align builtin structural tags with chat templates for reasoning and tool callsin v0.2.0
- day −93vLLM #41240bug[Bug]: DeepSeek V4 DSML tool parser mishandles wrapped and reserved argumentsfixed in v0.20.2
- day −93XGrammar v0.1.34supporttool-calling supportin v0.1.34
- day −93vLLM #41198fix[Bugfix] DSV32/V4 add missing type conversion for non-streaming tool callsin v0.20.1
- day −93TensorRT-LLM #13608support[None][feat] tool template and parser for dsv4in v1.3.0rc14
- day −94vLLM #41122bug[Bug] DeepSeek V4 tool calling returns boolean parameters as quoted strings which causes subagent failed in ClaudeCodefixed in v0.20.1
- day −94vLLM #41132bug[Bug]: DeepSeek V3.2 & V4 incorrect structured output when thinking enabledfixed in v0.20.1
- day −94XGrammar #596supportfeat: support DeepSeek_v4's tool calling format.in v0.1.34
- day −94SGLang #23915fixFix dsv4 self-closing invoke tags + accept reasoning_effort=maxin v0.5.11
- day −95vLLM v0.20.0supportmodel supportin v0.20.0
- day −95vLLM #40860support[Feat] DeepSeek V4 Rebased (#40860)in v0.20.0
- day −96vLLM #40806fix[Bugfix] Fix the DSML token leakage in DSV4/3.2in v0.20.0
- day −98vLLM #40800bug[Bug]: DeepSeek V4 intermittently leaks DSML fragments in auto + streaming mode, causing unstable tool-call parsingfixed in v0.20.0
- day −98vLLM #40801bug[Bug]: Title: DeepSeek V4 intermittently leaks DSML fragments in auto + streaming mode, causing unstable tool-call parsingfixed in v0.24.0
5open tool-calling issues upstreamvLLM 2 · SGLang 3 · TensorRT-LLM 0 · XGrammar 0 · oldest open 29 days
- 015 tool-calling bugs are open upstream for GLM 5.3 (vLLM 2, SGLang 3), the oldest for 29 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now5oldest 29 days
Opened / fixed5 / 0fixed: in a release
First server support—no support change found
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (15)
- day 73vLLM #62424support[Rust Frontend] Add native GLM-5.3 / GLM-5.3-Flash chat renderingmerged, not released
- day 73TensorRT-LLM #21109fix[None][fix] GLM reasoning parser: drop the owed </think> after a tool call; read the thinking mode off the promptmerged, not released
- day 73vLLM #64615fix[Bugfix][Frontend] Leave GLM-4.7 array and object tool arguments unpinnedmerged, not released
- day 72SGLang #47167bug[Bug] Tools + JSON-schema format cannot be combined: Kimi-K3 returns 400, glm47 / DeepSeek silently drop tool calls and invent the answer (breaks Codex --output-schema)open
- day 72vLLM #64535fix[Bugfix][Parser] Parse closed GLM-4.7 tool arguments once while streamingmerged, not released
- day 69SGLang #45659fix[Fix] rust-server: Match Python's chat layer on templates, streaming and lowered-chat errorsmerged, not released
- day 66SGLang #45953bug[Bug] Glm47MoeDetector.detect_and_parse drops tool calls when the function name is followed by a newline (non-streaming)open
- day 65vLLM #63185bug[Bug]: Responses API rejects `include: ["web_search_call.results", "web_search_call.action.sources"]` with 400, although the `web_search` tool is acceptedopen
- day 63SGLang #41233fix[Fix] glm47: Disable tool grammar for `tool_choice="none"`merged, not released
- day 63SGLang #45154feature[Feat] glm-5.3: Render GLM-5.3 and GLM-5.3-Flash chat prompts with a Python chat templatemerged, not released
- day 59vLLM #62051bug[Bug]: DeepSeek V4.1 tool parser drops malformed DSML blocks and emits invalid/empty args on truncationopen
- day 59vLLM #61574fix[Bugfix][Structured Output] Mask grammar-invalid implicit reasoning endsin v0.32.0
- day 57SGLang #43625support[sgl-router] Render GLM-5.3 and GLM-5.3-Flash through sglang-processorin v0.5.22
- day 45SGLang #42269bugresponse_format + tools on glm47: tool calls silently dropped, model invents a JSON answeropen
- day 35SGLang #40497fix[Docs] GLM-5.3/5.3-Flash cookbooks: enable reasoning/tool-call parsers by default via autoin v0.5.21
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now0
Opened / fixed0 / 0fixed: in a release
First server support—no support change found
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (1)
- day 8vLLM #63492fix[Bugfix][MiniMax-M3] Scope prompt reasoning-end checks to the last messagemerged, not released
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now0
Opened / fixed0 / 0fixed: in a release
First server supportday 5vLLM
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (5)
- day 86vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 78vLLM #57340fix[Rust Frontend] Build full-output grammars from initialized reasoning parsersin v0.31.0
- day 46vLLM #53054support[Rust Frontend] Add HY3 unified parser and local XGrammar structural-tag builderin v0.28.0
- day 5vLLM v0.25.0supportmodel supportin v0.25.0
- day −5vLLM #47192feature[Model] Support Hy3 token suffix and JSON Schema array typesin v0.25.0
19open tool-calling issues upstreamvLLM 10 · SGLang 7 · TensorRT-LLM 2 · XGrammar 0 · oldest open 115 days
- 0119 tool-calling bugs are open upstream for DeepSeek V4 Flash 0423 (vLLM 10, SGLang 7, TensorRT-LLM 2), the oldest for 115 days.
- 02SGLang v0.5.22 fixed #33163 ([Bug] deepseek-v4-flash toolcall error runner_backend from marlin to flashinfer_mxfp4) 12 days ago: providers on an older version still have the bug.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix12 daysreport to release
Open now19oldest 115 days
Opened / fixed40 / 21fixed: in a release
First server supportday 3vLLM; XGrammar day 5; SGLang day 22; TensorRT-LLM day 27
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (105)
- day 189TensorRT-LLM #21072fix[None][fix] Responses: detokenize for the parsers as chat completions doesmerged, not released
- day 188vLLM #58780fix[Bugfix][Parser] Stream DeepSeek V4/V3.2 string tool arguments incrementallymerged, not released
- day 188SGLang #46780fix[Feat] rust: Parse DeepSeek-V4/V4.1 output with deepseek-recipemerged, not released
- day 187vLLM #64326bug[Bug]: DelegatingParser.parse ignores model_output_token_ids, so markers written as plain text become tool calls or end reasoning (non-streaming)open
- day 186TensorRT-LLM #20873fix[None][fix] DeepSeek DSML tool parsers: the layout newlines around tool calls are not contentin v1.4.0rc2
- day 186vLLM #63762fix[Rust Frontend] Separate DeepSeek V4 content blocks like deepseek-recipemerged, not released
- day 186TensorRT-LLM #21025fix[None][fix] DeepSeek DSML tool parsers: keep string parameter values as writtenin v1.4.0rc2
- day 186TensorRT-LLM #21051fix[None][fix] DeepSeek renderers: tools join the leading system message; V4.1 renders developer as systemin v1.4.0rc2
- day 186SGLang #46706fix[Fix] dsv4: Render null contents, trailing tool-call turns and numeric string efforts with deepseek-recipemerged, not released
- day 185TensorRT-LLM #20989bug[Performance]: DeepSeek streaming tool parsers rescan the whole tool call on every increment (30 s of CPU for one 64k-character call)open
- day 184SGLang #46348bug[Bug] dsv41: deepseek-recipe rendering differs from the checkpoint's encoding.py (tool namespaces, current-turn reasoning before tool results, action)open
- day 183SGLang #45315feature[Feat] openai: Render DeepSeek-V4/V4.1 chat prompts with deepseek-recipemerged, not released
- day 183SGLang #45907support[Feat] openai: parse DeepSeek output through deepseek-recipemerged, not released
- day 183SGLang #46219feature[Feat] openai: Parse DeepSeek text for the native endpoints through deepseek-recipemerged, not released
- day 182vLLM #63314fix[Rust Frontend] Render empty tool call arguments as an empty object like Pythonmerged, not released
- day 181vLLM #62875fix[Rust Frontend] Match DeepSeek DSML markers by token identitymerged, not released
- day 180vLLM #62741feature[Rust Frontend] Support --exclude-tools-when-tool-choice-nonemerged, not released
- day 177SGLang #44941bugDeepSeek V3.2/V4/V4.1 streaming tool-call parser: pre-existing bugs found while fuzzingopen
- day 173SGLang #43625support[sgl-router] Render GLM-5.3 and GLM-5.3-Flash through sglang-processorin v0.5.22
- day 173SGLang #44243fix[Fix] sglang-processor: Follow Python's OpenAI layer through #43027, #37577, #43006, #43026, #43005in v0.5.22
- day 168SGLang #42999support[sgl-router] Parse tool calls when serving chat through /generatein v0.5.22
- day 167vLLM #60722bug[RFC]: Strcutured Outputs Not Respected When Reasoning Implicitly Endsfixed day 179, not released
- day 167vLLM #59627fix[Bugfix][Structured Output] Feed the token that implicitly ends reasoning to the grammarin v0.32.0
- day 167SGLang #42664fix[rust-processor] DeepSeek-V4 parity with SGLang, plus a reusable parity harness and skillin v0.5.22
- day 163vLLM #59926bug[Bug] DeepSeek-V4-Flash tool-call format failures associated with prefix-cache reuse; worst for short cached head + long uncached suffix; `cache_salt` reduces failures but is confounded with hit shape (v0.27.1, fp8 KV, MTP)open
- day 159vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 157SGLang #41226fix[sgl-router] Scope input_ids forwarding by renderer: all text chats for DeepSeek-V4in v0.5.21
- day 156XGrammar #936fixfix: bound DeepSeek V4.1 auto tool callsmerged, not released
- day 155SGLang #41317bug[Bug] DeepSeek V3.2/V4 DSML detector strips whitespace (e.g. trailing newline) from string="true" parameter valuesopen
- day 152vLLM #58280bug[Bug]: DeepSeek-V4 chat template puts `tools` into a synthetic leading system message instead of on the existing system message (diverges from the reference encoder)fixed in v0.31.0
- day 151SGLang #39632fixfix(function_call): buffer complete DeepSeek DSML invokesin v0.5.21
- day 149vLLM #57826bug[Bug]: minimax_m2 / glm47_moe / deepseek_v4 arg converters drop the last parameter when the model omits the closing tag (same class as #57699)open
- day 147SGLang #40236bug[Bug] DeepSeek V4/V3.2: DSML tool calls returned as content with no tool_calls (bare invoke, unterminated section, malformed sibling)open
- day 142SGLang #39277bug[Bug] b12x-vision dev-v4f-2dgx regresses DSV4 history argument encoding from #28035open
- day 140vLLM #56482bug[Bug]: DSML tool-call recovery (#55954) emits ghost empty-argument calls on truncation and narration, and recovers undeclared toolsopen
- day 140vLLM #56260fix[Bugfix][Rust Frontend][Renderer] Align DeepSeek tool-call arguments with deepseek-recipein v0.30.0
- day 139vLLM #56141fix[Bugfix] Tolerate misspelled DSML tool_calls wrapperin v0.30.0
- day 138vLLM #55954fix[Bugfix] Parse DSML tool calls when the model omits the tool_calls wrapperin v0.29.0
- day 137vLLM #55223fix[Perf] Eliminate full-history reasoning scans for structured outputsin v0.29.0
- day 131vLLM #54838fix[Bugfix] Implicitly close DeepSeek DSML parametersin v0.29.0
- day 130vLLM #53281fix[Bugfix][Rust Frontend] Fix adjacent DeepSeek V4 user content renderingin v0.29.0
- day 124vLLM #53831bug[Bug]: DeepSeek-V4-Flash tool_choice=auto silently drops tool calls at deep reasoning_effort (xhigh/max) — distinct from #41132/#51914fixed in v0.28.0
- day 124vLLM #48922fix[Bugfix] Guard tool call argument JSON parsing in chat message postprocessingin v0.29.0
- day 119vLLM #53046fix[Bugfix][Structured Output] Avoid spurious FSM errors after speculative reasoning endin v0.28.0
- day 118SGLang #35736bug[Bug] DeepSeek-V4 leaks raw DSML when tool_choice="none"open
- day 117vLLM #52846bug[Bug]: DeepSeek V4 buffers long string tool arguments until </parameter>fixed day 188, not released
- day 117SGLang #35563bug[Bug] deepseekv32/deepseekv4 tool-call parsers drop calls when the streamed output has two chunksfixed in v0.5.18
- day 117TensorRT-LLM #17903fix[#17580][fix] Stream the text that precedes a tool call in DeepSeek parsersin v1.3.0rc25
- day 116TensorRT-LLM #17573fix[#17572][fix] Emit the withheld buffer in DeepSeek streaming tool parsersin v1.3.0rc25
- day 110vLLM #51914bug[Bug] DeepSeek-V4-Flash-0731 intermittently emits malformed DSML tool-call start wrapper on v0.27.1 + DSparkopen
- day 110TensorRT-LLM #17572bug[Bug]: DeepSeek streaming tool parsers drop text they withheld while waiting on a tool-call tokenfixed in v1.3.0rc25
- day 110TensorRT-LLM #17580bug[Bug]: DeepSeek streaming tool parsers drop text that precedes a tool call in the same deltafixed in v1.3.0rc25
- day 110SGLang #34458fix[Fix] Make DeepSeek-V4 reasoning and tool-call streaming parsing chunk-invariantin v0.5.18
- day 109vLLM #51829bug[Bug]: DeepSeek V4 Python renderer misplaces tools when a system message existsfixed in v0.30.0
- day 109vLLM #51144feature[Rust Frontend] Support dynamic tools from developer messagesin v0.27.1
- day 108SGLang #34214bug[Bug] Streaming response truncates content when both content and tool_calls are present (SGLang 0.5.17)fixed day 183, not released
- day 108vLLM #51296fix[Bugfix] Align deepseek v4 parser thinking default with tokenizerin v0.27.0
- day 108vLLM #51727fix[Bugfix] Fix DeepSeek V4/3.2 tokenizer vocab size overcount crashing guided decodingin v0.27.1
- day 106vLLM #51467bug[Bug]: DeepSeek-V4-Flash-0731 `response_format` (structured output) crashes the vLLM EngineCore — `apply_grammar_bitmask` tensor size mismatch (4040 vs 4041)fixed in v0.27.1
- day 104SGLang #33867bug[Bug] /v1/responses API fails with 400 when function_call_output output is an arrayfixed in v0.5.22
- day 101vLLM #50924bug[Bug]: EngineCore dies on first guided-decoding request when dspark speculative decoding is enabled (grammar bitmask width mismatch)fixed in v0.27.1
- day 99SGLang #33163bug[Bug] deepseek-v4-flash toolcall error runner_backend from marlin to flashinfer_mxfp4fixed in v0.5.22
- day 96TensorRT-LLM #17021bugDeepSeek-V4: OpenAI `tools` field reorders the system prompt after the tool schemasopen
- day 92SGLang #31753fix[DSPARK] Grammar-constrained decoding, incl. tool_choice=autoin v0.5.17
- day 85vLLM #49002bug[Bug]: Speculative Decoding + Structured Output（tool call）组合下，decode 阶段出现秒级卡顿fixed in v0.30.0
- day 84vLLM #48931bug[Bug] DeepSeek-V4 tool-call parser leaks raw DSML into content when the model omits the `<｜DSML｜tool_calls>` START token (long context)fixed in v0.29.0
- day 77vLLM #48207bug[Bug]: Analysis for an Automated Request Validation and Normalization Module at the Upstream of vLLM (Hard Interception) to Fundamentally Solve DSML Leakage, Garbled Output, and Empty Responsesopen
- day 77SGLang #29883fix[BUG] fix strip streaming empty-string suffix from DSV4 tool argumentsin v0.5.15
- day 76vLLM #48089bug[Bug]: DeepSeek V4 reasoning/tool parsers produce unbalanced tags under concurrent loadfixed in v0.29.0
- day 76SGLang #30623other[Refactor] Share chat encoding dispatch between serving and offline toolsin v0.5.15
- day 75vLLM #47974bug[Bug]: vllm 校验arguments合法性open
- day 75vLLM #47986bug[Bug]: DeepSeek V4/V32 parser can unwrap arguments using the wrong tool's schema when parallel tool calls share identical raw argument textopen
- day 71vLLM #45877fix[Frontend] [Parser] Port DeepSeek V4 to streaming parser engine frameworkin v0.25.0
- day 65TensorRT-LLM #15414support[None][feat] DSv4: model, tokenizer, and integration coveragein v1.3.0rc20
- day 61XGrammar #671bug[Bug] deepseek_xml structural tag: regex pattern starting with an empty alternative (e.g. ^$|...) produces invalid EBNF (Expect element, but got |)fixed in v0.2.4
- day 60SGLang #28449supportExpand parser auto detection coveragein v0.5.14
- day 35SGLang #25676supportUpgrade xgrammar to 0.2.1in v0.5.13
- day 34vLLM #42879fix[Bugfix] Stream DeepSeek DSML tool-call argument deltas incrementallyin v0.22.0
- day 33vLLM #43582fix[Rust Frontend] Add reasoning/tool parser & renderer roundtrip testsin v0.22.0
- day 33vLLM #43662fix[Rust Frontend] Align tool parser fallback behavior between streaming & non-streaming pathsin v0.22.0
- day 29SGLang #26164fix[docs] DeepSeek-V4 cookbook: balanced MegaMoE cap, H200 Pro FP4 mem-frac, nsa-* compat, PD-disagg fixesin v0.5.12.post1
- day 27TensorRT-LLM v1.3.0rc15supportmodel supportin v1.3.0rc15
- day 27vLLM #43255other[CI] Add composed-schema regression tests for DeepSeek V3.2/V4 parsersin v0.22.0
- day 26SGLang #24251fix[RL][TITO] Preserve whitespace in reasoning parser outputsin v0.5.12.post1
- day 23vLLM #42878bug[Bug]: DeepSeek V4 DSML tool calls fake-stream arguments instead of incremental deltasfixed in v0.22.0
- day 23XGrammar #638fixfix(structural_tag): align deepseek_v3_2/v4 invoke separator with chat templatein v0.2.1
- day 22SGLang v0.5.12supporttool-calling supportin v0.5.12
- day 20XGrammar #637bugdeepseek_v3_2 structural tag: double-newline between invoke blocks blocks parallel tool calls at greedy decodingfixed in v0.2.1
- day 15SGLang #24793support[DSV4] Cherry pick missing commits from deepseek_v4 branch and enhance tests (#24793)in v0.5.12
- day 14SGLang #23882supportDeepseek V4 (#23882)in v0.5.12
- day 12vLLM #41801fix[Bugfix] DeepSeekV32/v4: respect string='true|false' attribute andunwrap arguments/input wrapperin v0.20.2
- day 7XGrammar #610fixfix: align builtin structural tags with chat templates for reasoning and tool callsin v0.2.0
- day 5vLLM #41240bug[Bug]: DeepSeek V4 DSML tool parser mishandles wrapped and reserved argumentsfixed in v0.20.2
- day 5XGrammar v0.1.34supporttool-calling supportin v0.1.34
- day 5vLLM #41198fix[Bugfix] DSV32/V4 add missing type conversion for non-streaming tool callsin v0.20.1
- day 5TensorRT-LLM #13608support[None][feat] tool template and parser for dsv4in v1.3.0rc14
- day 4vLLM #41122bug[Bug] DeepSeek V4 tool calling returns boolean parameters as quoted strings which causes subagent failed in ClaudeCodefixed in v0.20.1
- day 4vLLM #41132bug[Bug]: DeepSeek V3.2 & V4 incorrect structured output when thinking enabledfixed in v0.20.1
- day 4XGrammar #596supportfeat: support DeepSeek_v4's tool calling format.in v0.1.34
- day 4SGLang #23915fixFix dsv4 self-closing invoke tags + accept reasoning_effort=maxin v0.5.11
- day 3vLLM v0.20.0supportmodel supportin v0.20.0
- day 3vLLM #40860support[Feat] DeepSeek V4 Rebased (#40860)in v0.20.0
- day 2vLLM #40806fix[Bugfix] Fix the DSML token leakage in DSV4/3.2in v0.20.0
- day 0vLLM #40800bug[Bug]: DeepSeek V4 intermittently leaks DSML fragments in auto + streaming mode, causing unstable tool-call parsingfixed in v0.20.0
- day 0vLLM #40801bug[Bug]: Title: DeepSeek V4 intermittently leaks DSML fragments in auto + streaming mode, causing unstable tool-call parsingfixed in v0.24.0
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now0
Opened / fixed0 / 0fixed: in a release
First server supportday 3XGrammar; SGLang day 11
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (6)
- day 11SGLang v0.5.21supportmodel supportin v0.5.21
- day 9vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 8vLLM #58019support[Frontend] Support strict MiMo-V2.6 tool callingin v0.31.0
- day 3XGrammar v0.2.8supporttool-calling supportin v0.2.8
- day 1vLLM #57933support[Rust Frontend] Add MiMo V2.6 parser supportin v0.30.0
- day 1XGrammar #916supportfeat: support MiMo-V2.6 Pro and Flash structural tagsin v0.2.8
13open tool-calling issues upstreamvLLM 5 · SGLang 7 · TensorRT-LLM 1 · XGrammar 0 · oldest open 92 days
- 0113 tool-calling bugs are open upstream for Kimi K3 (vLLM 5, SGLang 7, TensorRT-LLM 1), the oldest for 92 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix11 daysreport to release
Open now13oldest 92 days
Opened / fixed15 / 2fixed: in a release
First server supportday 23SGLang; vLLM day 25; TensorRT-LLM day 27; XGrammar day 55
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (59)
- day 106TensorRT-LLM #21072fix[None][fix] Responses: detokenize for the parsers as chat completions doesmerged, not released
- day 105SGLang #47167bug[Bug] Tools + JSON-schema format cannot be combined: Kimi-K3 returns 400, glm47 / DeepSeek silently drop tool calls and invent the answer (breaks Codex --output-schema)open
- day 103vLLM #63941bug[Bug]: Kimi K3 non-streaming content leaks truncated XTML marker fragmentsopen
- day 102TensorRT-LLM #20993bug[Bug]: Kimi K3 and MiniMax-M3 tool parsers raise on deeply nested or very long numeric JSON argumentsopen
- day 102SGLang #45659fix[Fix] rust-server: Match Python's chat layer on templates, streaming and lowered-chat errorsmerged, not released
- day 102TensorRT-LLM #20775fix[TRTLLM-16711][fix] Remove the legacy renderer pathsin v1.4.0rc2
- day 101vLLM #63552other[Doc] Kimi-K3: InferenceX AgentX lanes as tuned presets; DCP only in presetsmerged, not released
- day 99vLLM #63062fix[Rust Frontend] End prefilled Kimi K3 reasoning at the response opener like Pythonmerged, not released
- day 98SGLang #45697bug[Bug] Kimi K3 tool-call parser (kimi_k3) does not stream tool-call arguments: whole call is buffered until <|close|>call, leaving the stream silent for minutes on long tool callsopen
- day 98vLLM #62507fix[Bugfix][Frontend][Kimi K3] Fix Kimi K3 XTML marker leakage in responses APImerged, not released
- day 97TensorRT-LLM #19964feature[TRTLLM-17242][feat] merge two renders into a standalone serve/render for dynamoin v1.4.0rc1
- day 97vLLM #62741feature[Rust Frontend] Support --exclude-tools-when-tool-choice-nonemerged, not released
- day 90SGLang #43608support[Feat] sgl-router: Render Kimi-K3 through sglang-processorin v0.5.22
- day 88vLLM #59190support[Rust Frontend] Build the Kimi K3 full-output grammar from its parser modein v0.32.0
- day 88vLLM #60781fix[Rust Frontend] Compose answer constraints into parser-owned output grammarsin v0.32.0
- day 84SGLang #43118bug[Bug] Kimi-K3 strict tool grammar: a free-text string argument can close itself early and smuggle in arguments the schema forbidsopen
- day 83vLLM #59408fix[Rust Frontend] Share `SchemaRoot` between argument coercion and grammarsin v0.32.0
- day 77vLLM #58358fix[Rust Frontend] Token-aware marker parsing for Kimi K3in v0.31.0
- day 77vLLM #59395fix[Rust Frontend] Split argument grammars into schema resolution and per-model renderingin v0.31.0
- day 76TensorRT-LLM #19499feature[None][feat] Add a per-model serving-extension registry to the OpenAI serverin v1.4.0rc0
- day 76vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 68vLLM #55269other[Rust Frontend] Introduce parser-owned output grammar interfacesin v0.31.0
- day 67XGrammar #912featurefeat: support string exclusions in JSON schema structural tagsin v0.2.8
- day 67XGrammar #915fixfeat: exclude control markers from Kimi-K3 tool-argument stringsin v0.2.8
- day 64vLLM #57272fix[Frontend] Upgrade XGrammar to 0.2.7 and Rust structural tags to 0.3.0in v0.30.0
- day 61SGLang #39583bug[Bug] Kimi K3 leaks raw XTML tool calls into content when no tools are declaredopen
- day 61SGLang #39640bug[Bug] SGLANG_FORWARD_UNKNOWN_TOOLS is not honoured by four tool-call detectors on the one-shot pathopen
- day 60XGrammar #889fixfix: correlate DeepSeek V4.1 parameter attributes with value typesin v0.2.7
- day 60XGrammar #892featurefeat: add a parallel_tool_calls param for builtin structural tagsin v0.2.7
- day 55SGLang #38587bug[Bug] Kimi-K3 strict tool-call grammar: additionalProperties dilutes a named property's type constraintopen
- day 55XGrammar v0.2.6supporttool-calling supportin v0.2.6
- day 53SGLang #37743fix[Kimi-K3] Recover the reply when the model skips the think channelin v0.5.20
- day 47SGLang #37430bug[Bug] Kimi-K3 tool_choice=required requests hang until ReadTimeout in 2P2D TP8/DCP8 deploymentopen
- day 43vLLM #54273bug[Bug]: Kimi-K3 parsers miss channel terminators - tool calls silently dropped, message terminator leaked into streamed contentopen
- day 43TensorRT-LLM #17845feature[TRTLLM-14764][feat] trtllm-serve: Kimi K3 API compliance for the Kimi Vendor Verifier (KVV)in v1.3.0rc25
- day 40TensorRT-LLM #17980fix[TRTLLM-15176][fix] Harden Kimi K3 tool-call parsingin v1.3.0rc25
- day 36vLLM #53246bug[Bug]: Kimi K3 tool calls returned empty tool calls when the model omits the think transition markeropen
- day 36vLLM #52889fix[Frontend] Prevent Kimi K3 reserved markers in response text (#52889)in v0.28.0
- day 33SGLang #34881fixStop losing Kimi-K3 tool calls to reasoning, constraint conflicts, and truncationin v0.5.18
- day 33SGLang #35399fix[Cherry-pick to release/v0.5.18] Stop losing Kimi-K3 tool calls to reasoning, constraint conflicts, and truncation (#34881)in v0.5.18
- day 29SGLang #34778fix[Fix] Work around xgrammar 0.2.1 negative integer minimum in Kimi-K3 structural tagsin v0.5.18
- day 27SGLang #34604bug[Bug] Kimi-K3 tool call parser fails ~8x/hour in production: TypeError 'string indices must be integers' and json 'unexpected character'fixed in v0.5.18
- day 27TensorRT-LLM v1.3.0rc24supporttool-calling supportin v1.3.0rc24
- day 26vLLM #51144feature[Rust Frontend] Support dynamic tools from developer messagesin v0.27.1
- day 25vLLM v0.27.0supportmodel supportin v0.27.0
- day 25XGrammar #806supportfeat: support kimi-k3 tool call stylein v0.2.6rc1
- day 23SGLang v0.5.17supporttool-calling supportin v0.5.17
- day 23TensorRT-LLM #17327support[TRTLLM-14814][feat] Kimi K3 serving parsers, chat template, and speculative decoding (suffix automaton + DFlash scaffold)in v1.3.0rc24
- day 22vLLM #51399bug[Bug]: Kimi K3 tool parser streams the raw <|close|>message<|sep|> marker into contentopen
- day 16SGLang #33025support[Kimi K3] Add reasoning, tool-call, and OpenAI serving supportin v0.5.17
- day 16SGLang #33183fixHarden Kimi K3 parser edge casesin v0.5.17
- day 16SGLang #33192fix[Kimi K3] Keep native tool-call format when constraints failin v0.5.17
- day 15vLLM #50604bug[Bug] Kimi-K3: tool-call arguments don't stream — entire payload emitted in one SSE delta after minutes of silenceopen
- day 15vLLM #50420fix[Frontend][Bugfix] Use default tool call IDs for Kimi K3 for conversation-level uniquenessin v0.27.0
- day 13vLLM #50295bug[Bug]: [Parser] Kimi-K3 tool-call IDs are message-scoped and repeat across turnsfixed in v0.27.0
- day 13vLLM #50093support[Model] Add Kimi K3 support: Python frontend [2/2]in v0.27.0
- day 13TensorRT-LLM #16972fix[None][fix] Preserve compact Kimi-K3 special tokensin v1.3.0rc23
- day 13SGLang #32757fixFix Kimi K3 reasoning leak in the Responses APIin v0.5.17
- day 12SGLang #32617support[Kimi K3] Support parser auto-detectionin v0.5.17
3open tool-calling issues upstreamvLLM 3 · SGLang 0 · TensorRT-LLM 0 · XGrammar 0 · oldest open 103 days
- 013 tool-calling bugs are open upstream for GLM 5.2 (vLLM 3), the oldest for 103 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix39 daysreport to release
Open now3oldest 103 days
Opened / fixed9 / 6fixed: in a release
First server supportday 10SGLang; vLLM day 111
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (17)
- day 136TensorRT-LLM #21109fix[None][fix] GLM reasoning parser: drop the owed </think> after a tool call; read the thinking mode off the promptmerged, not released
- day 115TensorRT-LLM #19935fix[None][fix] GLM tool and reasoning parsers: streaming matches the whole-text parsein v1.4.0rc0
- day 111vLLM v0.31.0supportmodel supportin v0.31.0
- day 106vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 63vLLM #51426fix[Rust Frontend] Fix GLM-5.2 chat template rendering parityin v0.28.0
- day 45SGLang #32522fix[Fix]: render tool_reference schema regardless of tool_result part orderin v0.5.17
- day 44vLLM #50399bug[Bug]: tool_choice: "required" causes tool_calls to continuously output until the maximum token is reached with GLM-5.2 on vLLM 0.26.0open
- day 41vLLM #49981bug[Bug]: tool_choice: "required" causes xgrammar FSM crash / infinite hang with GLM-5.2-NVFP4 on vLLM 0.26.0open
- day 34vLLM #49248bug[Bug]: glm47 tool parser silently returns empty arguments when the model omits the opening <arg_value> tagopen
- day 26SGLang #30925bug[Bug] GLM tool-call parser leaks raw <tool_call> markup into content when tool_choice="none"fixed in v0.5.15.post1
- day 23vLLM #48095bug[Bug]: GLM 5.2 forced toolcall left unparsed in content (glm 47 tool parser)fixed in v0.29.0
- day 17vLLM #47504bug[Bug]: GLM tool_call required + streaming causes JSON repetitionfixed in v0.31.0
- day 16SGLang #29914bug[Bug] [GLM5.2] When the tools are set to required mode, the output becomes repetitive and garbled.fixed in v0.5.19
- day 10SGLang v0.5.14supportmodel supportin v0.5.14
- day 8SGLang #29110bug[Bug] PD-disaggregation + EAGLE + grammar (tool_choice/structured): grammar token double-accept → FINISH_ABORTfixed in v0.5.15
- day 2vLLM #46040bug[Bug][ROCm]: GLM-5.2 emits `<tool_call>` inside `<think>` tagfixed in v0.24.0
- day 2vLLM #45915support[Frontend] Add Streaming Parser Engine and new GLM4.7/GLM5.1/GLM5.2 Parserin v0.24.0
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now0
Opened / fixed0 / 0fixed: in a release
First server supportday 3XGrammar; SGLang day 11
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (5)
- day 11SGLang v0.5.21supportmodel supportin v0.5.21
- day 9vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 8vLLM #58019support[Frontend] Support strict MiMo-V2.6 tool callingin v0.31.0
- day 3XGrammar v0.2.8supporttool-calling supportin v0.2.8
- day 1XGrammar #916supportfeat: support MiMo-V2.6 Pro and Flash structural tagsin v0.2.8
13open tool-calling issues upstreamvLLM 7 · SGLang 4 · TensorRT-LLM 1 · XGrammar 1 · oldest open 72 days
- 0113 tool-calling bugs are open upstream for Qwen3.8 27B (vLLM 7, SGLang 4, TensorRT-LLM 1, XGrammar 1), the oldest for 72 days.
- 02SGLang v0.5.22 fixed #34927 ([Bug] [Responses API] input_image parts in function_call_output are not converted to image_url: 400 ChatCompletionRequest validation error (post-#25881 builds), silently dropped on current main) 12 days ago: providers on an older version still have the bug.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix28 daysreport to release
Open now13oldest 72 days
Opened / fixed20 / 7fixed: in a release
First server supportday 8SGLang
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (24)
- day 74XGrammar #1019bug[Bug] `qwen_xml`: a string parameter with a `pattern` still accepts `</parameter>` inside its value with #989, so a second tool call fits in itopen
- day 73SGLang #45659fix[Fix] rust-server: Match Python's chat layer on templates, streaming and lowered-chat errorsmerged, not released
- day 61SGLang #43689support[Feat] sgl-router: Render Qwen3.8 (Max, 27B, Flash-Next) through sglang-processorin v0.5.22
- day 59vLLM #60136fix[Bugfix][Rust Frontend] Alias `qwen3_xml` to `qwen3_coder` and add `qwen3_json` for the Qwen JSON formatin v0.32.0
- day 57SGLang #43514bug[Bug] DFlash + Qwen3.8-27B: speculative decoding accept rate drops to 0 during servingopen
- day 53SGLang #42764bug[Bug] Qwen3CoderDetector: a string parameter whose value is `null` becomes null, and an invalid boolean becomes falseopen
- day 50vLLM #59834bug[Bug]: Responses API streaming regenerates output item id/call_id in response.completed (breaks strict clients)fixed in v0.31.0
- day 49SGLang #42260bug[Bug] Anthropic streaming drops a tool call's closing "}" when whitespace text arrives between its arguments (multiple tool calls, qwen3_coder)open
- day 42vLLM #58647bug[RFC]: Hardening /v1/messages (Anthropic API) for Claude Code across model familiesopen
- day 40vLLM #58227bug[Bug]: an unclosed <tool_call> opener written in prose holds the Qwen3 parser in tool state and swallows a later genuine tool callfixed in v0.32.0
- day 39vLLM #58147bug[Bug]: Qwen3 parser turns tool-call markup the model quotes into real tool calls (names outside request.tools, and duplicates of offered ones)open
- day 36vLLM #57699bug[Bug]: qwen3_coder tool parser drops the last parameter when the model omits </parameter> before </function> (non-streaming {}, streaming unterminated JSON)open
- day 26vLLM #56077bug[Bug]: ngram speculative decoding corrupts qwen3_coder tool-call parser output (empty/mangled arguments) for Qwen3.x modelsfixed in v0.32.0
- day 21vLLM #55284bug[Bug]: Responses API streaming aborts mid-stream (nameless ResponseFunctionToolCallItem) with Qwen3.8 + qwen3_coder + DFlash speculative decodingopen
- day 21vLLM #55313bug[Bug]: Qwen3.8 + DSpark + streaming json_schema can desync XGrammar and emit 250k trailing spaces with HTTP 200open
- day 18vLLM #54808bug[Bug]: qwen3_coder / qwen3_xml parser silently ignores tool_choice "required" and named function on /v1/chat/completions (0.28.0)fixed in v0.31.0
- day 11vLLM #53777bug[Bug] DFlash2 spec decode + xgrammar: deterministic "Failed to advance FSM" on json_object grammar (same draft token every time)fixed in v0.30.0
- day 8vLLM #53359bug[Bug]: tool_call_parser_invocations_total no longer increments for engine-backed parsers (regression after ParserEngine refactor)open
- day 8SGLang v0.5.18supportmodel supportin v0.5.18
- day 7vLLM #53181bug[Bug] Xgrammar "Failed to advance FSM" still fires after #52805 (0.27.2rc1.dev256+geac636a7f) — follow-up to #52852open
- day 7TensorRT-LLM #18084bugMODEL_TYPE_TO_TOOL_PARSER maps qwen3_5/qwen3_5_moe to a parser whose format the template does not emitopen
- day 6SGLang #35692bug[Bug] Anthropic endpoint: tool_reference parts in tool_result break chat templates without deferred-reference support (HTTP 500)open
- day 5vLLM #52852bug[Bug] Streaming completions end with no finish_reason / no [DONE] — xgrammar "Failed to advance FSM" is one logged trigger, but some drops are fully silent (no engine error)fixed in v0.28.0
- day 1SGLang #34927bug[Bug] [Responses API] input_image parts in function_call_output are not converted to image_url: 400 ChatCompletionRequest validation error (post-#25881 builds), silently dropped on current mainfixed in v0.5.22
14open tool-calling issues upstreamvLLM 7 · SGLang 5 · TensorRT-LLM 2 · XGrammar 0 · oldest open 115 days
- 0114 tool-calling bugs are open upstream for DeepSeek V4 Pro 0813 (vLLM 7, SGLang 5, TensorRT-LLM 2), the oldest for 115 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix19 daysreport to release
Open now14oldest 115 days
Opened / fixed27 / 13fixed: in a release
First server supportday −107vLLM; XGrammar day −105; SGLang day −88; TensorRT-LLM day −83
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (83)
- day 79TensorRT-LLM #21072fix[None][fix] Responses: detokenize for the parsers as chat completions doesmerged, not released
- day 78vLLM #58780fix[Bugfix][Parser] Stream DeepSeek V4/V3.2 string tool arguments incrementallymerged, not released
- day 78SGLang #46780fix[Feat] rust: Parse DeepSeek-V4/V4.1 output with deepseek-recipemerged, not released
- day 77vLLM #64326bug[Bug]: DelegatingParser.parse ignores model_output_token_ids, so markers written as plain text become tool calls or end reasoning (non-streaming)open
- day 76TensorRT-LLM #20873fix[None][fix] DeepSeek DSML tool parsers: the layout newlines around tool calls are not contentin v1.4.0rc2
- day 76vLLM #63762fix[Rust Frontend] Separate DeepSeek V4 content blocks like deepseek-recipemerged, not released
- day 76TensorRT-LLM #21025fix[None][fix] DeepSeek DSML tool parsers: keep string parameter values as writtenin v1.4.0rc2
- day 76TensorRT-LLM #21051fix[None][fix] DeepSeek renderers: tools join the leading system message; V4.1 renders developer as systemin v1.4.0rc2
- day 76SGLang #46706fix[Fix] dsv4: Render null contents, trailing tool-call turns and numeric string efforts with deepseek-recipemerged, not released
- day 75TensorRT-LLM #20989bug[Performance]: DeepSeek streaming tool parsers rescan the whole tool call on every increment (30 s of CPU for one 64k-character call)open
- day 74SGLang #46348bug[Bug] dsv41: deepseek-recipe rendering differs from the checkpoint's encoding.py (tool namespaces, current-turn reasoning before tool results, action)open
- day 73SGLang #45315feature[Feat] openai: Render DeepSeek-V4/V4.1 chat prompts with deepseek-recipemerged, not released
- day 73SGLang #45907support[Feat] openai: parse DeepSeek output through deepseek-recipemerged, not released
- day 73SGLang #46219feature[Feat] openai: Parse DeepSeek text for the native endpoints through deepseek-recipemerged, not released
- day 72vLLM #63314fix[Rust Frontend] Render empty tool call arguments as an empty object like Pythonmerged, not released
- day 70vLLM #62741feature[Rust Frontend] Support --exclude-tools-when-tool-choice-nonemerged, not released
- day 67SGLang #44941bugDeepSeek V3.2/V4/V4.1 streaming tool-call parser: pre-existing bugs found while fuzzingopen
- day 63SGLang #43625support[sgl-router] Render GLM-5.3 and GLM-5.3-Flash through sglang-processorin v0.5.22
- day 63SGLang #44243fix[Fix] sglang-processor: Follow Python's OpenAI layer through #43027, #37577, #43006, #43026, #43005in v0.5.22
- day 58SGLang #42999support[sgl-router] Parse tool calls when serving chat through /generatein v0.5.22
- day 57vLLM #60722bug[RFC]: Strcutured Outputs Not Respected When Reasoning Implicitly Endsfixed day 69, not released
- day 57vLLM #59627fix[Bugfix][Structured Output] Feed the token that implicitly ends reasoning to the grammarin v0.32.0
- day 49vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 47SGLang #41226fix[sgl-router] Scope input_ids forwarding by renderer: all text chats for DeepSeek-V4in v0.5.21
- day 46XGrammar #936fixfix: bound DeepSeek V4.1 auto tool callsmerged, not released
- day 45SGLang #41317bug[Bug] DeepSeek V3.2/V4 DSML detector strips whitespace (e.g. trailing newline) from string="true" parameter valuesopen
- day 41SGLang #39632fixfix(function_call): buffer complete DeepSeek DSML invokesin v0.5.21
- day 39vLLM #57826bug[Bug]: minimax_m2 / glm47_moe / deepseek_v4 arg converters drop the last parameter when the model omits the closing tag (same class as #57699)open
- day 37SGLang #40236bug[Bug] DeepSeek V4/V3.2: DSML tool calls returned as content with no tool_calls (bare invoke, unterminated section, malformed sibling)open
- day 32SGLang #39277bug[Bug] b12x-vision dev-v4f-2dgx regresses DSV4 history argument encoding from #28035open
- day 30vLLM #56260fix[Bugfix][Rust Frontend][Renderer] Align DeepSeek tool-call arguments with deepseek-recipein v0.30.0
- day 28vLLM #55954fix[Bugfix] Parse DSML tool calls when the model omits the tool_calls wrapperin v0.29.0
- day 27vLLM #55223fix[Perf] Eliminate full-history reasoning scans for structured outputsin v0.29.0
- day 20vLLM #53281fix[Bugfix][Rust Frontend] Fix adjacent DeepSeek V4 user content renderingin v0.29.0
- day 9vLLM #53227bug[Bug]: DeepSeek V4 streaming tool calls leak DSML markup into argumentsfixed in v0.29.0
- day 7vLLM #52846bug[Bug]: DeepSeek V4 buffers long string tool arguments until </parameter>fixed day 78, not released
- day 7SGLang #35563bug[Bug] deepseekv32/deepseekv4 tool-call parsers drop calls when the streamed output has two chunksfixed in v0.5.18
- day 7TensorRT-LLM #17903fix[#17580][fix] Stream the text that precedes a tool call in DeepSeek parsersin v1.3.0rc25
- day 6TensorRT-LLM #17573fix[#17572][fix] Emit the withheld buffer in DeepSeek streaming tool parsersin v1.3.0rc25
- day 0TensorRT-LLM #17572bug[Bug]: DeepSeek streaming tool parsers drop text they withheld while waiting on a tool-call tokenfixed in v1.3.0rc25
- day 0TensorRT-LLM #17580bug[Bug]: DeepSeek streaming tool parsers drop text that precedes a tool call in the same deltafixed in v1.3.0rc25
- day 0SGLang #34458fix[Fix] Make DeepSeek-V4 reasoning and tool-call streaming parsing chunk-invariantin v0.5.18
- day −1vLLM #51144feature[Rust Frontend] Support dynamic tools from developer messagesin v0.27.1
- day −2vLLM #51296fix[Bugfix] Align deepseek v4 parser thinking default with tokenizerin v0.27.0
- day −14TensorRT-LLM #17021bugDeepSeek-V4: OpenAI `tools` field reorders the system prompt after the tool schemasopen
- day −18SGLang #31753fix[DSPARK] Grammar-constrained decoding, incl. tool_choice=autoin v0.5.17
- day −25vLLM #49002bug[Bug]: Speculative Decoding + Structured Output（tool call）组合下，decode 阶段出现秒级卡顿fixed in v0.30.0
- day −33vLLM #48207bug[Bug]: Analysis for an Automated Request Validation and Normalization Module at the Upstream of vLLM (Hard Interception) to Fundamentally Solve DSML Leakage, Garbled Output, and Empty Responsesopen
- day −33SGLang #29883fix[BUG] fix strip streaming empty-string suffix from DSV4 tool argumentsin v0.5.15
- day −34vLLM #48089bug[Bug]: DeepSeek V4 reasoning/tool parsers produce unbalanced tags under concurrent loadfixed in v0.29.0
- day −34SGLang #30623other[Refactor] Share chat encoding dispatch between serving and offline toolsin v0.5.15
- day −35vLLM #47974bug[Bug]: vllm 校验arguments合法性open
- day −35vLLM #47986bug[Bug]: DeepSeek V4/V32 parser can unwrap arguments using the wrong tool's schema when parallel tool calls share identical raw argument textopen
- day −45TensorRT-LLM #15414support[None][feat] DSv4: model, tokenizer, and integration coveragein v1.3.0rc20
- day −49XGrammar #671bug[Bug] deepseek_xml structural tag: regex pattern starting with an empty alternative (e.g. ^$|...) produces invalid EBNF (Expect element, but got |)fixed in v0.2.4
- day −50SGLang #28449supportExpand parser auto detection coveragein v0.5.14
- day −75SGLang #25676supportUpgrade xgrammar to 0.2.1in v0.5.13
- day −76vLLM #42879fix[Bugfix] Stream DeepSeek DSML tool-call argument deltas incrementallyin v0.22.0
- day −77vLLM #43582fix[Rust Frontend] Add reasoning/tool parser & renderer roundtrip testsin v0.22.0
- day −77vLLM #43662fix[Rust Frontend] Align tool parser fallback behavior between streaming & non-streaming pathsin v0.22.0
- day −81SGLang #26164fix[docs] DeepSeek-V4 cookbook: balanced MegaMoE cap, H200 Pro FP4 mem-frac, nsa-* compat, PD-disagg fixesin v0.5.12.post1
- day −83TensorRT-LLM v1.3.0rc15supportmodel supportin v1.3.0rc15
- day −83vLLM #43255other[CI] Add composed-schema regression tests for DeepSeek V3.2/V4 parsersin v0.22.0
- day −87vLLM #42878bug[Bug]: DeepSeek V4 DSML tool calls fake-stream arguments instead of incremental deltasfixed in v0.22.0
- day −87XGrammar #638fixfix(structural_tag): align deepseek_v3_2/v4 invoke separator with chat templatein v0.2.1
- day −88SGLang v0.5.12supporttool-calling supportin v0.5.12
- day −90XGrammar #637bugdeepseek_v3_2 structural tag: double-newline between invoke blocks blocks parallel tool calls at greedy decodingfixed in v0.2.1
- day −95SGLang #24793support[DSV4] Cherry pick missing commits from deepseek_v4 branch and enhance tests (#24793)in v0.5.12
- day −96SGLang #23882supportDeepseek V4 (#23882)in v0.5.12
- day −98vLLM #41801fix[Bugfix] DeepSeekV32/v4: respect string='true|false' attribute andunwrap arguments/input wrapperin v0.20.2
- day −103XGrammar #610fixfix: align builtin structural tags with chat templates for reasoning and tool callsin v0.2.0
- day −105vLLM #41240bug[Bug]: DeepSeek V4 DSML tool parser mishandles wrapped and reserved argumentsfixed in v0.20.2
- day −105XGrammar v0.1.34supporttool-calling supportin v0.1.34
- day −105vLLM #41198fix[Bugfix] DSV32/V4 add missing type conversion for non-streaming tool callsin v0.20.1
- day −105TensorRT-LLM #13608support[None][feat] tool template and parser for dsv4in v1.3.0rc14
- day −106vLLM #41132bug[Bug]: DeepSeek V3.2 & V4 incorrect structured output when thinking enabledfixed in v0.20.1
- day −106XGrammar #596supportfeat: support DeepSeek_v4's tool calling format.in v0.1.34
- day −106SGLang #23915fixFix dsv4 self-closing invoke tags + accept reasoning_effort=maxin v0.5.11
- day −107vLLM v0.20.0supportmodel supportin v0.20.0
- day −107vLLM #40860support[Feat] DeepSeek V4 Rebased (#40860)in v0.20.0
- day −108vLLM #40806fix[Bugfix] Fix the DSML token leakage in DSV4/3.2in v0.20.0
- day −110vLLM #40800bug[Bug]: DeepSeek V4 intermittently leaks DSML fragments in auto + streaming mode, causing unstable tool-call parsingfixed in v0.20.0
- day −110vLLM #40801bug[Bug]: Title: DeepSeek V4 intermittently leaks DSML fragments in auto + streaming mode, causing unstable tool-call parsingfixed in v0.24.0
7open tool-calling issues upstreamvLLM 5 · SGLang 1 · TensorRT-LLM 1 · XGrammar 0 · oldest open 115 days
- 017 tool-calling bugs are open upstream for MiniMax M3 (vLLM 5, SGLang 1, TensorRT-LLM 1), the oldest for 115 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now7oldest 115 days
Opened / fixed7 / 0fixed: in a release
First server supportday 23TensorRT-LLM; vLLM day 29; SGLang day 69; XGrammar day 107
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (24)
- day 148TensorRT-LLM #20993bug[Bug]: Kimi K3 and MiniMax-M3 tool parsers raise on deeply nested or very long numeric JSON argumentsopen
- day 146SGLang #46133fix[Fix] minimax-m3: Type nested tool-call arguments and count self-opened reasoningmerged, not released
- day 129vLLM #59408fix[Rust Frontend] Share `SchemaRoot` between argument coercion and grammarsin v0.32.0
- day 128vLLM #59743fix[Frontend] Port the MiniMax M3 tool parser to the parser enginein v0.32.0
- day 121vLLM #59004fix[Rust Frontend] Relax schema-aware tool argument conversionin v0.31.0
- day 119SGLang #40930fixAdd MiniMax arch fallback to auto parser resolutionin v0.5.21
- day 117vLLM #58647bug[RFC]: Hardening /v1/messages (Anthropic API) for Claude Code across model familiesopen
- day 107XGrammar v0.2.7supporttool-calling supportin v0.2.7
- day 102XGrammar #713supportfeat: support minimax-m3 style tool call structural tagin v0.2.7
- day 89SGLang #36626fix[Fix] Resolve tool argument types through top-level anyOf/oneOf/allOfin v0.5.19
- day 78vLLM #52613bug[Bug]: MiniMax-M3 (MXFP8) streaming: generation stops (empty delta, finish_reason=stop) right after reasoning closes, whenever a tool call should follow — non-streaming with identical payload succeedsopen
- day 72vLLM #51235fix[Rust Frontend] Upgrade MiniJinja to 2.22 & remove method lookup workaroundin v0.27.1
- day 70vLLM #51549bug[Bug]: RustToolParser streaming can drop buffered text at stream endopen
- day 69SGLang v0.5.17supportmodel supportin v0.5.17
- day 65vLLM #51073bug[Bug]: MinimaxM3ToolParser (Rust) fails on MiniMaxAI/MiniMax-M3-MXFP8 tool output; raw markup + ']<]minimax[>[' special token leaks into contentopen
- day 54SGLang #32286bug[Bug] MiniMax-M3 tool-call parser: top-level `oneOf` parameters schema not resolved (numbers become strings, arrays leak raw tags)fixed day 153, not released
- day 41SGLang #28715support[minimax-m3] Split 4/4: model + VL + glue + function-call + fp8 quant + generic infrain v0.5.15.post1
- day 38vLLM #47998bug[Bug] MiniMax-M3 tool-call parser buffers arguments — no incremental streamingopen
- day 32TensorRT-LLM #15813fix[https://nvbugs/6385673][fix] Add `isinstance` guards for `parameters`, `properties`, and the per-parameter…in v1.3.0rc21
- day 29vLLM v0.24.0supportmodel supportin v0.24.0
- day 23TensorRT-LLM v1.3.0rc19supportmodel supportin v1.3.0rc19
- day 23vLLM #45718fix[Bugfix] Parse MiniMax M3 streaming reasoning by text markersin v0.24.0
- day 19TensorRT-LLM #15292support[None][feat] BREAKING: Add MiniMax-M3 PyTorch backend bring-up with API changes (#15292)in v1.3.0rc19
- day 18vLLM #45826fix[Rust Frontend][Perf] O(n) argument scan in tool parserin v0.24.0
14open tool-calling issues upstreamvLLM 7 · SGLang 5 · TensorRT-LLM 2 · XGrammar 0 · oldest open 115 days
- 0114 tool-calling bugs are open upstream for DeepSeek V4 Pro 0423 (vLLM 7, SGLang 5, TensorRT-LLM 2), the oldest for 115 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix19 daysreport to release
Open now14oldest 115 days
Opened / fixed27 / 13fixed: in a release
First server supportday 3vLLM; XGrammar day 5; SGLang day 22; TensorRT-LLM day 27
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (83)
- day 189TensorRT-LLM #21072fix[None][fix] Responses: detokenize for the parsers as chat completions doesmerged, not released
- day 188vLLM #58780fix[Bugfix][Parser] Stream DeepSeek V4/V3.2 string tool arguments incrementallymerged, not released
- day 188SGLang #46780fix[Feat] rust: Parse DeepSeek-V4/V4.1 output with deepseek-recipemerged, not released
- day 187vLLM #64326bug[Bug]: DelegatingParser.parse ignores model_output_token_ids, so markers written as plain text become tool calls or end reasoning (non-streaming)open
- day 186TensorRT-LLM #20873fix[None][fix] DeepSeek DSML tool parsers: the layout newlines around tool calls are not contentin v1.4.0rc2
- day 186vLLM #63762fix[Rust Frontend] Separate DeepSeek V4 content blocks like deepseek-recipemerged, not released
- day 186TensorRT-LLM #21025fix[None][fix] DeepSeek DSML tool parsers: keep string parameter values as writtenin v1.4.0rc2
- day 186TensorRT-LLM #21051fix[None][fix] DeepSeek renderers: tools join the leading system message; V4.1 renders developer as systemin v1.4.0rc2
- day 186SGLang #46706fix[Fix] dsv4: Render null contents, trailing tool-call turns and numeric string efforts with deepseek-recipemerged, not released
- day 185TensorRT-LLM #20989bug[Performance]: DeepSeek streaming tool parsers rescan the whole tool call on every increment (30 s of CPU for one 64k-character call)open
- day 184SGLang #46348bug[Bug] dsv41: deepseek-recipe rendering differs from the checkpoint's encoding.py (tool namespaces, current-turn reasoning before tool results, action)open
- day 183SGLang #45315feature[Feat] openai: Render DeepSeek-V4/V4.1 chat prompts with deepseek-recipemerged, not released
- day 183SGLang #45907support[Feat] openai: parse DeepSeek output through deepseek-recipemerged, not released
- day 183SGLang #46219feature[Feat] openai: Parse DeepSeek text for the native endpoints through deepseek-recipemerged, not released
- day 182vLLM #63314fix[Rust Frontend] Render empty tool call arguments as an empty object like Pythonmerged, not released
- day 180vLLM #62741feature[Rust Frontend] Support --exclude-tools-when-tool-choice-nonemerged, not released
- day 177SGLang #44941bugDeepSeek V3.2/V4/V4.1 streaming tool-call parser: pre-existing bugs found while fuzzingopen
- day 173SGLang #43625support[sgl-router] Render GLM-5.3 and GLM-5.3-Flash through sglang-processorin v0.5.22
- day 173SGLang #44243fix[Fix] sglang-processor: Follow Python's OpenAI layer through #43027, #37577, #43006, #43026, #43005in v0.5.22
- day 168SGLang #42999support[sgl-router] Parse tool calls when serving chat through /generatein v0.5.22
- day 167vLLM #60722bug[RFC]: Strcutured Outputs Not Respected When Reasoning Implicitly Endsfixed day 179, not released
- day 167vLLM #59627fix[Bugfix][Structured Output] Feed the token that implicitly ends reasoning to the grammarin v0.32.0
- day 159vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 157SGLang #41226fix[sgl-router] Scope input_ids forwarding by renderer: all text chats for DeepSeek-V4in v0.5.21
- day 156XGrammar #936fixfix: bound DeepSeek V4.1 auto tool callsmerged, not released
- day 155SGLang #41317bug[Bug] DeepSeek V3.2/V4 DSML detector strips whitespace (e.g. trailing newline) from string="true" parameter valuesopen
- day 151SGLang #39632fixfix(function_call): buffer complete DeepSeek DSML invokesin v0.5.21
- day 149vLLM #57826bug[Bug]: minimax_m2 / glm47_moe / deepseek_v4 arg converters drop the last parameter when the model omits the closing tag (same class as #57699)open
- day 147SGLang #40236bug[Bug] DeepSeek V4/V3.2: DSML tool calls returned as content with no tool_calls (bare invoke, unterminated section, malformed sibling)open
- day 142SGLang #39277bug[Bug] b12x-vision dev-v4f-2dgx regresses DSV4 history argument encoding from #28035open
- day 140vLLM #56260fix[Bugfix][Rust Frontend][Renderer] Align DeepSeek tool-call arguments with deepseek-recipein v0.30.0
- day 138vLLM #55954fix[Bugfix] Parse DSML tool calls when the model omits the tool_calls wrapperin v0.29.0
- day 137vLLM #55223fix[Perf] Eliminate full-history reasoning scans for structured outputsin v0.29.0
- day 130vLLM #53281fix[Bugfix][Rust Frontend] Fix adjacent DeepSeek V4 user content renderingin v0.29.0
- day 119vLLM #53227bug[Bug]: DeepSeek V4 streaming tool calls leak DSML markup into argumentsfixed in v0.29.0
- day 117vLLM #52846bug[Bug]: DeepSeek V4 buffers long string tool arguments until </parameter>fixed day 188, not released
- day 117SGLang #35563bug[Bug] deepseekv32/deepseekv4 tool-call parsers drop calls when the streamed output has two chunksfixed in v0.5.18
- day 117TensorRT-LLM #17903fix[#17580][fix] Stream the text that precedes a tool call in DeepSeek parsersin v1.3.0rc25
- day 116TensorRT-LLM #17573fix[#17572][fix] Emit the withheld buffer in DeepSeek streaming tool parsersin v1.3.0rc25
- day 110TensorRT-LLM #17572bug[Bug]: DeepSeek streaming tool parsers drop text they withheld while waiting on a tool-call tokenfixed in v1.3.0rc25
- day 110TensorRT-LLM #17580bug[Bug]: DeepSeek streaming tool parsers drop text that precedes a tool call in the same deltafixed in v1.3.0rc25
- day 110SGLang #34458fix[Fix] Make DeepSeek-V4 reasoning and tool-call streaming parsing chunk-invariantin v0.5.18
- day 109vLLM #51144feature[Rust Frontend] Support dynamic tools from developer messagesin v0.27.1
- day 108vLLM #51296fix[Bugfix] Align deepseek v4 parser thinking default with tokenizerin v0.27.0
- day 96TensorRT-LLM #17021bugDeepSeek-V4: OpenAI `tools` field reorders the system prompt after the tool schemasopen
- day 92SGLang #31753fix[DSPARK] Grammar-constrained decoding, incl. tool_choice=autoin v0.5.17
- day 85vLLM #49002bug[Bug]: Speculative Decoding + Structured Output（tool call）组合下，decode 阶段出现秒级卡顿fixed in v0.30.0
- day 77vLLM #48207bug[Bug]: Analysis for an Automated Request Validation and Normalization Module at the Upstream of vLLM (Hard Interception) to Fundamentally Solve DSML Leakage, Garbled Output, and Empty Responsesopen
- day 77SGLang #29883fix[BUG] fix strip streaming empty-string suffix from DSV4 tool argumentsin v0.5.15
- day 76vLLM #48089bug[Bug]: DeepSeek V4 reasoning/tool parsers produce unbalanced tags under concurrent loadfixed in v0.29.0
- day 76SGLang #30623other[Refactor] Share chat encoding dispatch between serving and offline toolsin v0.5.15
- day 75vLLM #47974bug[Bug]: vllm 校验arguments合法性open
- day 75vLLM #47986bug[Bug]: DeepSeek V4/V32 parser can unwrap arguments using the wrong tool's schema when parallel tool calls share identical raw argument textopen
- day 65TensorRT-LLM #15414support[None][feat] DSv4: model, tokenizer, and integration coveragein v1.3.0rc20
- day 61XGrammar #671bug[Bug] deepseek_xml structural tag: regex pattern starting with an empty alternative (e.g. ^$|...) produces invalid EBNF (Expect element, but got |)fixed in v0.2.4
- day 60SGLang #28449supportExpand parser auto detection coveragein v0.5.14
- day 35SGLang #25676supportUpgrade xgrammar to 0.2.1in v0.5.13
- day 34vLLM #42879fix[Bugfix] Stream DeepSeek DSML tool-call argument deltas incrementallyin v0.22.0
- day 33vLLM #43582fix[Rust Frontend] Add reasoning/tool parser & renderer roundtrip testsin v0.22.0
- day 33vLLM #43662fix[Rust Frontend] Align tool parser fallback behavior between streaming & non-streaming pathsin v0.22.0
- day 29SGLang #26164fix[docs] DeepSeek-V4 cookbook: balanced MegaMoE cap, H200 Pro FP4 mem-frac, nsa-* compat, PD-disagg fixesin v0.5.12.post1
- day 27TensorRT-LLM v1.3.0rc15supportmodel supportin v1.3.0rc15
- day 27vLLM #43255other[CI] Add composed-schema regression tests for DeepSeek V3.2/V4 parsersin v0.22.0
- day 23vLLM #42878bug[Bug]: DeepSeek V4 DSML tool calls fake-stream arguments instead of incremental deltasfixed in v0.22.0
- day 23XGrammar #638fixfix(structural_tag): align deepseek_v3_2/v4 invoke separator with chat templatein v0.2.1
- day 22SGLang v0.5.12supporttool-calling supportin v0.5.12
- day 20XGrammar #637bugdeepseek_v3_2 structural tag: double-newline between invoke blocks blocks parallel tool calls at greedy decodingfixed in v0.2.1
- day 15SGLang #24793support[DSV4] Cherry pick missing commits from deepseek_v4 branch and enhance tests (#24793)in v0.5.12
- day 14SGLang #23882supportDeepseek V4 (#23882)in v0.5.12
- day 12vLLM #41801fix[Bugfix] DeepSeekV32/v4: respect string='true|false' attribute andunwrap arguments/input wrapperin v0.20.2
- day 7XGrammar #610fixfix: align builtin structural tags with chat templates for reasoning and tool callsin v0.2.0
- day 5vLLM #41240bug[Bug]: DeepSeek V4 DSML tool parser mishandles wrapped and reserved argumentsfixed in v0.20.2
- day 5XGrammar v0.1.34supporttool-calling supportin v0.1.34
- day 5vLLM #41198fix[Bugfix] DSV32/V4 add missing type conversion for non-streaming tool callsin v0.20.1
- day 5TensorRT-LLM #13608support[None][feat] tool template and parser for dsv4in v1.3.0rc14
- day 4vLLM #41132bug[Bug]: DeepSeek V3.2 & V4 incorrect structured output when thinking enabledfixed in v0.20.1
- day 4XGrammar #596supportfeat: support DeepSeek_v4's tool calling format.in v0.1.34
- day 4SGLang #23915fixFix dsv4 self-closing invoke tags + accept reasoning_effort=maxin v0.5.11
- day 3vLLM v0.20.0supportmodel supportin v0.20.0
- day 3vLLM #40860support[Feat] DeepSeek V4 Rebased (#40860)in v0.20.0
- day 2vLLM #40806fix[Bugfix] Fix the DSML token leakage in DSV4/3.2in v0.20.0
- day 0vLLM #40800bug[Bug]: DeepSeek V4 intermittently leaks DSML fragments in auto + streaming mode, causing unstable tool-call parsingfixed in v0.20.0
- day 0vLLM #40801bug[Bug]: Title: DeepSeek V4 intermittently leaks DSML fragments in auto + streaming mode, causing unstable tool-call parsingfixed in v0.24.0
15open tool-calling issues upstreamvLLM 6 · SGLang 6 · TensorRT-LLM 2 · XGrammar 1 · oldest open 347 days
- 0115 tool-calling bugs are open upstream for gpt-oss-20b (vLLM 6, SGLang 6, TensorRT-LLM 2, XGrammar 1), the oldest for 347 days.
- 02SGLang v0.5.22 fixed #31019 ([Bug] GPT-OSS + require_reasoning + json_schema emits malformed Harmony without Role and Channel) 12 days ago: providers on an older version still have the bug.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix35 daysreport to release
Open now15oldest 347 days
Opened / fixed36 / 21fixed: in a release
First server supportday 11TensorRT-LLM; vLLM day 13; SGLang day 18
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (90)
- day 450SGLang #47205bug[Bug] Reasoning grammar gate never opens when the think-end marker splits under split_special_tokensopen
- day 445vLLM #63345fix[Bugfix][Frontend] Apply batch tool_choice rules to streamed tool callsmerged, not released
- day 444vLLM #63346bug[Bug]: Streaming chat keeps tool calls for `tool_choice: "none"` (gpt-oss) and `tool_choice: null`fixed day 445, not released
- day 444vLLM #63148fix[Frontend] Share non-streaming chat message assembly with batch derendermerged, not released
- day 444vLLM #63314fix[Rust Frontend] Render empty tool call arguments as an empty object like Pythonmerged, not released
- day 443vLLM #62786fix[Bugfix][Frontend] Render GPT-OSS chats with Harmony in /tokenizemerged, not released
- day 442TensorRT-LLM #19964feature[TRTLLM-17242][feat] merge two renders into a standalone serve/render for dynamoin v1.4.0rc1
- day 438SGLang #44931bug[Bug] gpt-oss on /v1/chat/completions drops replayed reasoning: the template reads `thinking`, SGLang only forwards `reasoning_content`open
- day 429SGLang #43156bug[Bug] HarmonyParser silently deletes streamed content that is a prefix of "commentary"open
- day 427vLLM #54835fix[Bugfix][Frontend] Apply model-default reasoning parser in GPU-less render serverin v0.32.0
- day 425vLLM #59859fix[Bugfix][Responses API] Reuse streamed item ids in final harmony responsein v0.31.0
- day 422SGLang #42143bug[Bug] HarmonyParser streaming: arguments of a tool call on the analysis channel are emitted as reasoningopen
- day 422vLLM #59005support[Rust Frontend] Add `hf` parser for response templatesin v0.31.0
- day 421TensorRT-LLM #19499feature[None][feat] Add a per-model serving-extension registry to the OpenAI serverin v1.4.0rc0
- day 420SGLang #41747bug[Bug] gpt-oss: `tool_choice: "required"` returns raw harmony control tokens as assistant contentopen
- day 418vLLM #58930bug[Bug]: validate_xgrammar_grammar skips unsupported-feature checks for JSON schemas nested in structural tagsopen
- day 417SGLang #41315bug[Bug] gpt-oss detector misses tool calls that are the first message or lack <|constrain|>jsonopen
- day 417vLLM #58825bug[Bug]: gpt-oss (HarmonyParser): a stray 'commentary to=assistant' header is returned as a tool call named 'assistant<|channel|>analysis'open
- day 409vLLM #57593bug[Feature]: Share non-streaming chat message assembly between `OpenAIServingChat` and batch derenderfixed day 444, not released
- day 406TensorRT-LLM #19223bug[Bug]: standalone Responses API input, streaming, and tool continuation failures with Agentic APIopen
- day 387vLLM #52222fix[Bugfix][GPT-OSS] Fix strict tool-call grammar to accept Harmony rendersin v0.29.0
- day 361vLLM #45560feature[GPT-OSS] Strict tool call and constrained decoding for Harmonyin v0.27.0
- day 351vLLM #49493bug[Bug]: Harmony models reject namespace tool type in Responses APIopen
- day 347XGrammar #701bug[Bug] Stop token that is also a grammar terminal can never be acceptedfixed day 448, not released
- day 342SGLang #31019bug[Bug] GPT-OSS + require_reasoning + json_schema emits malformed Harmony without Role and Channelfixed in v0.5.22
- day 329vLLM #35076fix[Bugfix] Propagate default stop_token_ids to per-request SamplingParamsin v0.25.0
- day 328vLLM #46800support[Rust Frontend] Add Harmony Renderer for GPT-OSSin v0.24.0
- day 327vLLM #46940bug[Bug]: Responses API — Harmony model function_call.status is null instead of "completed"fixed in v0.25.0
- day 322vLLM #45048fix[Bugfix] GPT-OSS Autodrop reasoning in Response API and cleanupin v0.24.0
- day 314vLLM #45464fix[Bugfix] Chat Completions Harmony Refactor Clean upin v0.24.0
- day 311vLLM #45104fix[Refactor] Chat Completions Streaming Harmony Refactor and Bugfixesin v0.23.0
- day 281vLLM #42454fix[Bugfix] Handle real-world gpt-oss tool call output in Harmony parsingin v0.21.0
- day 275SGLang #23952featurefeat(reasoning): auto-detect reasoning/tool-call parser from chat templatein v0.5.12
- day 275SGLang #23953featurefeat(constrained): two-phase reasoning grammar + --enable-strict-thinkingin v0.5.12
- day 267XGrammar #607fixfix: fix gpt-oss's tool-calling format.in v0.2.0
- day 237TensorRT-LLM #12517fix[#11992][fix] Support include_stop_token_in_output in gRPC request managerin v1.3.0rc10
- day 219vLLM #36849bug[Bug]: GPT-OSS-20B /v1/chat/completions streaming crashes with tool_call_parser=openai (IndexError in chat_completion_stream_generator)fixed in v0.18.1
- day 218vLLM #36560feature[CI] Add bfcl tool call correctness evalin v0.18.0
- day 192vLLM #34454fix[Bugfix]: Fix structured output in multi-turn gpt-ossin v0.16.0
- day 176vLLM #33210bug[Bug]: gpt-oss chat format mismatch with HF apply_chat_templatefixed in v0.23.0
- day 174vLLM #33094bug[CI Failure]: Entrypoints Test (Responses)fixed in v0.15.0
- day 169vLLM #32751bug[CI Failure]: mi325_1 Entrypoints Integration Test (Responses API)fixed in v0.16.0
- day 169vLLM #32791bug[Bug]: chat.completions returns content: null for GPT-OSS multi-turn with json_objectfixed in v0.16.0
- day 167vLLM #32587bug[Bug]: gpt-oss special tokens leak into tool names and responses despite `reasoning_parser: "openai_gptoss"`open
- day 162TensorRT-LLM #10651bug[Bug]: Missing Stop Tokens in responses_utils.py Causes Content Leakage in Tool Use Scenariosfixed in v1.3.0rc18
- day 162vLLM #26291fix[MODEL] Fix handling of multiple channels for gpt-oss with speculative decoding in v0.14.0
- day 157vLLM #31761feature[Frontend] Add MCP tool streaming support to Responses APIin v0.14.0
- day 138vLLM #30608fix[BugFix]fix gpt-oss v1/completions response bug (#30608)in v0.14.0
- day 136vLLM #28139bug[Frontend][Bug] allow tool calls in analysis channelin v0.14.0
- day 136vLLM #30867fix[Bugfix] Fix tool_choice="none" being ignored by GPT-OSS/harmony modelsin v0.13.0
- day 134vLLM #30854bug[CI Failure]: Entrypoints Integration Test (API Server)fixed in v0.13.0
- day 132vLLM #30648fix[Bugfix] Drop empty tool_calls lists to keep assistant replies in chat templatein v0.13.0
- day 123vLLM #26237other[Tests] Tool call tests for openai/gpt-oss-20bin v0.13.0
- day 113vLLM #29541bug[CI Failure]: mi325_1: Entrypoints Integration Test (API Server 1)fixed in v0.18.0
- day 105TensorRT-LLM #9256bug[Bug]: trtllm-serve: Harmony-format tokens and reasoning fields emitted in responses for gpt-oss-120Bopen
- day 100SGLang #13163otherRemove EBNF Composer (#13163)in gateway-v0.1.6
- day 94SGLang #12431feature[router] Support structured model output for openai and grpc routerin gateway-v0.1.6
- day 91vLLM #28041bug[Bug]: online serving GPT-OSS does not stop at <|call|> token during reasoning (v0.11.0)fixed in v0.11.1
- day 90SGLang #12570bug[Bug] grpc function call not workingfixed in v0.5.5
- day 88SGLang #12226fixForward unknown tool calls instead of dropping (#12226)in v0.5.5
- day 85vLLM #27675fix[Frontend] [gpt-oss] Tool json call parsing error retryin v0.11.1
- day 79SGLang #11944fixAllow to disable batch decoding.in v0.5.4
- day 74vLLM #25515support[GPT-OSS] Structure_Tag support for gpt-oss tool-call in cotin v0.11.1
- day 70vLLM #24317feature[Feature][Responses API] Stream Function Call - harmonyin v0.11.1
- day 68SGLang #10577fixFix the GPT function calling regex to allow dash in the name (#10577)in v0.5.4
- day 59vLLM #24768fix[openai] Fix missing tool usage check (system message)in v0.11.1
- day 58vLLM #26075bug[Bug][gpt-oss]: Chat completion no built-in tool result with streaming outputfixed in v0.11.1
- day 50vLLM #25560bug[Bug]: GPT oss show tool call argument in response contentfixed in v0.11.0
- day 48SGLang #10738bug[Bug] gpt oss tool parser not workingfixed in v0.5.6
- day 48vLLM #24628support[Frontend] Responses API MCP tools for built in tools and to pass through headersin v0.11.0
- day 41SGLang #9657fixfix: gpt-oss streaming dropping normal content when tools are provided but not usedin v0.5.3
- day 34vLLM #24409fix[gpt-oss][Responses API] Fix the function call id formatin v0.10.2
- day 29vLLM #24170bug[Bug]: Intermittent "Unknown recipient: None" when calling gpt-oss-20b with Responsesfixed in v0.13.0
- day 23TensorRT-LLM #7261support[TRTLLM-7207][feat] Chat completions API for gpt-ossin v1.1.0rc2
- day 20SGLang #9190fixFix Harmony reasoning parser for and auto-separation for gpt-oss models (#9190)in v0.5.2
- day 19vLLM #23409fix[gpt-oss] Streaming Output for Python Toolin v0.10.2
- day 18SGLang v0.5.1supporttool-calling supportin v0.5.1
- day 17TensorRT-LLM #7163bug[Bug]: With the gpt-oss-20b model in trtllm-serve, tool call information is not captured in the final messagefixed in v1.1.0rc4
- day 15SGLang #9368fixFix gpt-oss response api streaming issuein v0.5.1
- day 13vLLM v0.10.1supporttool-calling supportin v0.10.1
- day 13SGLang #9316fix[MISC] use dynamic choices for tool-call-parser argumentin v0.5.1
- day 11TensorRT-LLM v1.1.0rc0supportmodel supportin v1.1.0rc0
- day 9SGLang #9166bug[Bug] openai/gpt-oss-20b fails with tools enabled — Jinja template error ('dict object' has no attribute 'name')fixed in v0.5.1
- day 7SGLang #9102bug[Bug] gpt-oss function toolfixed in v0.5.1
- day 7SGLang #9043support(gpt-oss, oai, chat): Remove Harmony Integration and Implement Native GPT-OSS Tool Call Supportin v0.5.1
- day 7SGLang #9097fixbugfix: Fix the commentary msg extraction in GptOssDetectorin v0.5.1
- day 3vLLM #22519bug[Bug]: [gpt oss 20b] [tool_call] Unexpected token 12606 while expecting start token 200006fixed in v0.25.0
- day 3SGLang #8976bug[Bug] # Chat Completion API fails when Harmony models generate commentary channel messagesfixed in v0.5.4
- day 1SGLang #8837supportSupport v1/responses and use harmony in serving_chatin v0.5.1
- day 1vLLM #22374support[gpt-oss] Add loop for built-in tool callin v0.10.1rc1
5open tool-calling issues upstreamvLLM 4 · SGLang 0 · TensorRT-LLM 0 · XGrammar 1 · oldest open 213 days
- 015 tool-calling bugs are open upstream for Qwen3.5-9B (vLLM 4, XGrammar 1), the oldest for 213 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix9 daysreport to release
Open now5oldest 213 days
Opened / fixed9 / 4fixed: in a release
First server supportday −14SGLang; TensorRT-LLM day −14; vLLM day −3
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (24)
- day 211XGrammar #940feature[Python] Add opt-in token_markers to get_model_structural_tagmerged, not released
- day 204vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 203XGrammar #939bugBuiltin structural tags admit control markers assembled from ordinary sub-tokensfixed day 211, not released
- day 199vLLM #58647bug[RFC]: Hardening /v1/messages (Anthropic API) for Claude Code across model familiesopen
- day 188XGrammar #892featurefeat: add a parallel_tool_calls param for builtin structural tagsin v0.2.7
- day 179vLLM #55495bug[Bug]: qwen3_xml tool parser emits a truncated `arguments` string with a leaked `</parameter` tag; the Responses API then rejects every later request of the conversation with `400 Expecting value`open
- day 169TensorRT-LLM #17575fix[#17574][fix] Complete zero-argument tool calls in the streaming tool parserin v1.3.0rc25
- day 140TensorRT-LLM #16866fix[https://nvbugs/6240584][fix] Qwen3ToolParser: bare-JSON fallback for reasoning-preceded tool callsin v1.3.0rc23
- day 135vLLM #44993fix[Bugfix][Structured Output][Spec Decode] Advance grammar across reasoning boundaryin v0.26.0
- day 132vLLM #49205bug[Bug]: StreamingParserEngine leaks a bare tag fragment (e.g. '<tool_') into content when a PreLexedTerminal supersedes a buffered text-token prefix of the same tagopen
- day 122SGLang #30744bug[Bug] Qwen3.5-9B runtime LoRA: tool calls swallowed into reasoning_content / malformed (4.8% defective turns); absent for base+identity on same server and for merged weights on vLLM/hostedfixed in v0.5.20
- day 94vLLM #45003feature[Frontend] Support strict mode for tool callingin v0.23.0
- day 78vLLM #43582fix[Rust Frontend] Add reasoning/tool parser & renderer roundtrip testsin v0.22.0
- day 71vLLM #43220bug[Bug]: Streaming reasoning tokens truncated when `</think>` and `<tool_call>` appear in the same deltafixed in v0.22.0
- day 55XGrammar #625featurefeat: expose normalize_tool_choice and unify Qwen XML structural tag buildersin v0.2.1
- day 52XGrammar #610fixfix: align builtin structural tags with chat templates for reasoning and tool callsin v0.2.0
- day 50XGrammar #609fixrefactor: unify reasoning parameter and rename model keysin v0.2.0
- day 22vLLM #38696bug[Bug]: qwen3.5 when enable response_format json_schema outputs garbled spacesopen
- day 6vLLM #36774fix[Bugfix] fix Qwen3.5 tool calling bugin v0.18.0
- day 1vLLM #36769bug[Bug]: Qwen3.5 tool calling bugfixed in v0.18.0
- day 1vLLM #36794bug[Feature]: Improve Error Message When Max Length Reached With Tool Call=Requiredfixed in v0.18.0
- day −3vLLM v0.17.0supportmodel supportin v0.17.0
- day −14SGLang v0.5.9supportmodel supportin v0.5.9
- day −14TensorRT-LLM v1.3.0rc5supportmodel supportin v1.3.0rc5
14open tool-calling issues upstreamvLLM 7 · SGLang 4 · TensorRT-LLM 3 · XGrammar 0 · oldest open 347 days
- 0114 tool-calling bugs are open upstream for gpt-oss-120b (vLLM 7, SGLang 4, TensorRT-LLM 3), the oldest for 347 days.
- 02SGLang v0.5.22 fixed #31019 ([Bug] GPT-OSS + require_reasoning + json_schema emits malformed Harmony without Role and Channel) 12 days ago: providers on an older version still have the bug.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix24 daysreport to release
Open now14oldest 347 days
Opened / fixed36 / 22fixed: in a release
First server supportday 11TensorRT-LLM; vLLM day 13; SGLang day 18
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (84)
- day 450SGLang #47205bug[Bug] Reasoning grammar gate never opens when the think-end marker splits under split_special_tokensopen
- day 444vLLM #63148fix[Frontend] Share non-streaming chat message assembly with batch derendermerged, not released
- day 444vLLM #63314fix[Rust Frontend] Render empty tool call arguments as an empty object like Pythonmerged, not released
- day 442TensorRT-LLM #19964feature[TRTLLM-17242][feat] merge two renders into a standalone serve/render for dynamoin v1.4.0rc1
- day 438SGLang #44931bug[Bug] gpt-oss on /v1/chat/completions drops replayed reasoning: the template reads `thinking`, SGLang only forwards `reasoning_content`open
- day 429SGLang #43156bug[Bug] HarmonyParser silently deletes streamed content that is a prefix of "commentary"open
- day 422SGLang #42143bug[Bug] HarmonyParser streaming: arguments of a tool call on the analysis channel are emitted as reasoningopen
- day 422vLLM #59005support[Rust Frontend] Add `hf` parser for response templatesin v0.31.0
- day 421TensorRT-LLM #19499feature[None][feat] Add a per-model serving-extension registry to the OpenAI serverin v1.4.0rc0
- day 418vLLM #58930bug[Bug]: validate_xgrammar_grammar skips unsupported-feature checks for JSON schemas nested in structural tagsopen
- day 409vLLM #57593bug[Feature]: Share non-streaming chat message assembly between `OpenAIServingChat` and batch derenderfixed day 444, not released
- day 384vLLM #53626bugUnsupported tool_choice in Responses API (Harmony/gpt-oss) returns HTTP 501, should be 4xxopen
- day 384TensorRT-LLM #17464fix[https://nvbugs/6456085][fix] Harmony: stop discarding malformed tool-call messages silentlyin v1.3.0rc25
- day 372vLLM #51977bug[Bug]: openai_harmony.HarmonyError: unexpected tokens remaining in message header on v0.26.0 (gpt-oss-120b)open
- day 368TensorRT-LLM #17437bug[Bug]: Interior control-token rejection can split a tool-call prelude and silently drop tool_callsopen
- day 361vLLM #45560feature[GPT-OSS] Strict tool call and constrained decoding for Harmonyin v0.27.0
- day 351vLLM #49493bug[Bug]: Harmony models reject namespace tool type in Responses APIopen
- day 343TensorRT-LLM #16377bug[Bug]: Eagle3 speculative decoding silently drops tool_calls for gpt-oss on trtllm-serveopen
- day 342SGLang #31019bug[Bug] GPT-OSS + require_reasoning + json_schema emits malformed Harmony without Role and Channelfixed in v0.5.22
- day 329vLLM #47185fix[Refactor][GPT-OSS] Harmony Responses API Refactor to use HarmonyParserin v0.25.0
- day 328vLLM #46800support[Rust Frontend] Add Harmony Renderer for GPT-OSSin v0.24.0
- day 314vLLM #45464fix[Bugfix] Chat Completions Harmony Refactor Clean upin v0.24.0
- day 311vLLM #45104fix[Refactor] Chat Completions Streaming Harmony Refactor and Bugfixesin v0.23.0
- day 309vLLM #44686fixFix Harmony tool descriptions for optional fieldsin v0.23.0
- day 300vLLM #44216bug[Bug]: [gpt-oss-120b] tool_choice="required" ignored on /v1/chat/completions (v0.20.1)open
- day 289vLLM #43338bug[Bug] PR #36138 grammar-mask spec-decode fix doesn't handle multi-token reasoning boundaries (gpt-oss/openai_gptoss still bleeds; Qwen3 fixed)open
- day 281vLLM #42454fix[Bugfix] Handle real-world gpt-oss tool call output in Harmony parsingin v0.21.0
- day 275SGLang #23952featurefeat(reasoning): auto-detect reasoning/tool-call parser from chat templatein v0.5.12
- day 275SGLang #23953featurefeat(constrained): two-phase reasoning grammar + --enable-strict-thinkingin v0.5.12
- day 267XGrammar #607fixfix: fix gpt-oss's tool-calling format.in v0.2.0
- day 230vLLM #37937bug[Bug]: IndexError: prev_tool_call_arr list index out of range when streaming tool call hits max_tokens (openai parser)fixed in v0.18.1
- day 176vLLM #33210bug[Bug]: gpt-oss chat format mismatch with HF apply_chat_templatefixed in v0.23.0
- day 169vLLM #32791bug[Bug]: chat.completions returns content: null for GPT-OSS multi-turn with json_objectfixed in v0.16.0
- day 161TensorRT-LLM #10612bug[Bug]: Structured output constantly misses proper channel establishment for batch and streaming and result is often malformed.fixed in v1.3.0rc1
- day 157vLLM #30437fix[Bugfix] missing tokens occur in harmony streamingin v0.14.0
- day 157vLLM #31761feature[Frontend] Add MCP tool streaming support to Responses APIin v0.14.0
- day 138vLLM #30608fix[BugFix]fix gpt-oss v1/completions response bug (#30608)in v0.14.0
- day 136vLLM #28139bug[Frontend][Bug] allow tool calls in analysis channelin v0.14.0
- day 129vLLM #28729fix[Bugfix] Multiple fixes for gpt-oss Chat Completion promptingin v0.13.0
- day 122SGLang #14462bug[Bug] Data race when using spec_v2 in certain guided decoding pathsfixed in v0.5.9
- day 122vLLM #29972fix[Bugfix] Fix parse_output_message crash on commentary with no recipientin v0.13.0
- day 120vLLM #29998bug[Bug]: cannot send two POST to /v1/chat/completions endpoint with identic tool function name with model GPT-OSS-120Bfixed in v0.16.0
- day 114SGLang #14029bug[Bug] grammar backend paths yield empty results for spec decoding + reasoning_parser on Hopper/Blackwell for some modelsfixed in v0.5.6
- day 105TensorRT-LLM #9256bug[Bug]: trtllm-serve: Harmony-format tokens and reasoning fields emitted in responses for gpt-oss-120Bopen
- day 100SGLang #13163otherRemove EBNF Composer (#13163)in gateway-v0.1.6
- day 99SGLang #13106bug[Bug] Using SpecDecoding results in empty outputs for structured outputsfixed in v0.5.6
- day 94vLLM #28000fix[Bugfix] Fix and add tests for GptOss reasoning parserin v0.11.1
- day 93SGLang #12720bug[Bug] [Router] Function calls fail if mcp server is also providedfixed in v0.5.5
- day 92SGLang #12667bug[Bug] [Router] Responses API does not correctly generate tool calls in multitool scenariosfixed in v0.5.5
- day 92SGLang #12669bug[Bug] GRPC Router issue with MCP Tool handlingfixed in v0.5.5
- day 91vLLM #28041bug[Bug]: online serving GPT-OSS does not stop at <|call|> token during reasoning (v0.11.0)fixed in v0.11.1
- day 91SGLang #12652bug[Router] Responses API streaming format incompatibility with OpenAI API for function callsfixed in v0.5.5
- day 91SGLang #12653bug[Bug] Responses API streaming format incompatibility with OpenAI API for function callsfixed in v0.5.5
- day 90SGLang #12567bug[Bug] gpt-oss-120b + spec + overlap scheduler issues (garbage output & server failure on "tool_choice": "required")fixed in v0.5.8
- day 88SGLang #12226fixForward unknown tool calls instead of dropping (#12226)in v0.5.5
- day 86SGLang #12153support[router] Harmony Pipeline: Chat Completion & Responses API with MCP Supportin v0.5.5
- day 85vLLM #27675fix[Frontend] [gpt-oss] Tool json call parsing error retryin v0.11.1
- day 83vLLM #26962feature[gpt-oss][2/N] Support input_messages in responsesRequestin v0.11.1
- day 79SGLang #11944fixAllow to disable batch decoding.in v0.5.4
- day 77vLLM #27243bug[Bug]: [vLLM v0.11.0 + GPT-OSS-120B]: when using tool calling, Unexpected token 200012 while expecting start token 200006 error occurred with a probability of approximately 10–20%.fixed in v0.11.1
- day 74vLLM #25515support[GPT-OSS] Structure_Tag support for gpt-oss tool-call in cotin v0.11.1
- day 68SGLang #10577fixFix the GPT function calling regex to allow dash in the name (#10577)in v0.5.4
- day 59vLLM #24768fix[openai] Fix missing tool usage check (system message)in v0.11.1
- day 50vLLM #25514fix[Bugfix] [Frontend] Cleanup gpt-oss non-streaming chat tool callsin v0.11.0
- day 48SGLang #10738bug[Bug] gpt oss tool parser not workingfixed in v0.5.6
- day 41SGLang #9657fixfix: gpt-oss streaming dropping normal content when tools are provided but not usedin v0.5.3
- day 35vLLM #23386support[gpt-oss] Harmony changes with container tool supportin v0.10.2
- day 34vLLM #24409fix[gpt-oss][Responses API] Fix the function call id formatin v0.10.2
- day 31SGLang #10089bug[Bug] muti fc call doesn't work on GPT OSS 120bfixed in v0.5.6
- day 31vLLM #22386support[gpt-oss] tool parser supports for /chat/completions [1/n]in v0.10.2
- day 23TensorRT-LLM #7261support[TRTLLM-7207][feat] Chat completions API for gpt-ossin v1.1.0rc2
- day 20SGLang #9190fixFix Harmony reasoning parser for and auto-separation for gpt-oss models (#9190)in v0.5.2
- day 18SGLang v0.5.1supporttool-calling supportin v0.5.1
- day 15SGLang #9390bug[Bug] Can't use builtin browser tool for GPTOssfixed in v0.5.4
- day 15SGLang #9368fixFix gpt-oss response api streaming issuein v0.5.1
- day 13vLLM v0.10.1supporttool-calling supportin v0.10.1
- day 13SGLang #9316fix[MISC] use dynamic choices for tool-call-parser argumentin v0.5.1
- day 11TensorRT-LLM v1.1.0rc0supportmodel supportin v1.1.0rc0
- day 7SGLang #9102bug[Bug] gpt-oss function toolfixed in v0.5.1
- day 7SGLang #9043support(gpt-oss, oai, chat): Remove Harmony Integration and Implement Native GPT-OSS Tool Call Supportin v0.5.1
- day 7SGLang #9097fixbugfix: Fix the commentary msg extraction in GptOssDetectorin v0.5.1
- day 1vLLM #22337bug[Usage]: gpt-oss-120b tool callsfixed in v0.18.0
- day 1SGLang #8837supportSupport v1/responses and use harmony in serving_chatin v0.5.1
- day 1vLLM #22374support[gpt-oss] Add loop for built-in tool callin v0.10.1rc1
12open tool-calling issues upstreamvLLM 7 · SGLang 4 · TensorRT-LLM 1 · XGrammar 0 · oldest open 208 days
- 0112 tool-calling bugs are open upstream for Gemma 4 31B (vLLM 7, SGLang 4, TensorRT-LLM 1), the oldest for 208 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix23 daysreport to release
Open now12oldest 208 days
Opened / fixed37 / 25fixed: in a release
First server supportday 1vLLM; TensorRT-LLM day 15; XGrammar day 27; SGLang day 33
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (81)
- day 208vLLM #63768fix[Rust Frontend] Accept parenthesized and name-only Gemma 4 tool calls like Pythonmerged, not released
- day 207TensorRT-LLM #20995bug[Bug]: Gemma4 tool parser raises RecursionError on arguments nested about 1000 levels deepopen
- day 204vLLM #63059fix[Rust Frontend] Accept Gemma 4 tool calls without the call prefix like Pythonmerged, not released
- day 202vLLM #62741feature[Rust Frontend] Support --exclude-tools-when-tool-choice-nonemerged, not released
- day 195vLLM #61494fix[Bugfix][Parser] Stream Gemma4 string tool arguments incrementallyin v0.32.0
- day 189vLLM #59627fix[Bugfix][Structured Output] Feed the token that implicitly ends reasoning to the grammarin v0.32.0
- day 187SGLang #42791bug[Bug] gemma4 tool-call parser hangs forever on a stray "]" in an array argumentopen
- day 182vLLM #59005support[Rust Frontend] Add `hf` parser for response templatesin v0.31.0
- day 177SGLang #41316bug[Bug] Gemma4Detector returns null and exponent-notation numbers as stringsopen
- day 176vLLM #58647bug[RFC]: Hardening /v1/messages (Anthropic API) for Claude Code across model familiesopen
- day 175XGrammar #923supportfeat: support Gemma 4 tool calling with the <|"|> string delimitermerged, not released
- day 166SGLang #39640bug[Bug] SGLANG_FORWARD_UNKNOWN_TOOLS is not honoured by four tool-call detectors on the one-shot pathopen
- day 166vLLM #56200fix[Refactor] Derive is_reasoning_end from the engine grammarin v0.30.0
- day 160vLLM #55417fix[Rust Frontend] Correctly parse whitespace framing in model outputin v0.29.0
- day 153vLLM #54881bug[Bug]: Gemma 4 skips reasoning and repeats JSON when json_schema is enabledfixed in v0.29.0
- day 148vLLM #54256bug[Bug]: Gemma4 parser drops tool calls when transitioning out of reasoning channel (<channel|>call:...)open
- day 147TensorRT-LLM #14290feature[TRTLLM-12758][feat] honor named function in tool_choice for non-harmony modelsin v1.3.0rc25
- day 145vLLM #53657fix[Bugfix] Handle parenthesized Gemma4 tool callsin v0.28.0
- day 139SGLang #35564bug[Bug] 8 tool-call parsers give wrong or missing calls in token-by-token streamingfixed day 209, not released
- day 123vLLM #50927bug[Bug]: stack overflow causing system crash: Rust unbounded recursion in the Gemma4 unified parserfixed in v0.29.0
- day 119vLLM #50477bug[Bug]: gemma4 parser silently ignores named forced tool_choice on 0.26.0 (worked on 0.21.0)open
- day 116vLLM #49745other[Refactor] Remove dead code in multiple files (#49745)in v0.27.0
- day 111vLLM #49412bug[Bug]: Streaming vs non-streaming content whitespace mismatch at tool-call boundaries (ParserEngine: qwen3/nemotron_v3/seed_oss/glm47_moe/gemma4)open
- day 99vLLM #48217bug[Bug]: Gemma4 parser classifies plain output without channel markers differently in streaming and non-streaming modefixed in v0.26.0
- day 97SGLang #30556bug[Bug] N Options brokenfixed in v0.5.20
- day 96vLLM #47906bug[Bug]: Gemma4 engine-based tool parser returns all bare-value arguments as strings (scilent behavior change)open
- day 92vLLM #47289fix[Rust Frontend] [Parser]Recover buffered text from incomplete tool calls at EOSin v0.25.0
- day 91vLLM #47401bug[Bug]: ][Docker] vllm/vllm-openai-cpu:v0.24.0 image missing tool_chat_template_gemma4.jinja file in examplesfixed in v0.25.0
- day 89vLLM #46875fix[Parser][Bugfix] Ensure tool call or other special tokens don't leak in non-streaming tool parsingin v0.25.0
- day 84vLLM #46602other[Rust Frontend] Migrate gemma4 to unified parserin v0.24.0
- day 77vLLM #46091fix[Bugfix] [Parser] Fix empty tool block silently dropping subsequent contentin v0.24.0
- day 76vLLM #45832fix[Bugfix][Gemma4] Fix parsing when thinking is disabledin v0.24.0
- day 75vLLM #45834bug[Bug]: [Gemma4] Post-tool reasoning can leak after final tool response with enable_thinking=truefixed in v0.24.0
- day 75vLLM #45553fix[Bugfix][Gemma4] Fix offline parser truncation, adjust_request token leak, and chat template syncin v0.24.0
- day 75vLLM #45795fix[Bugfix] Gemma4: skip forced JSON for required/named tool choicein v0.24.0
- day 74vLLM #45588fix[Frontend] Replace legacy Gemma4 parsers with engine-based implementationin v0.24.0
- day 71vLLM #45449bug[Bug]: Gemma4 Tool Parser: Intermittently Missing `function.name` in Streaming Modefixed in v0.24.0
- day 65vLLM #44715bug[Bug]: gemma4_tool_parser._parse_gemma4_args does not strip <|"|> from dict-key positionsfixed in v0.24.0
- day 63vLLM #44522bugGemma-4 tool call parser leaks raw tokens (<|" and "|>) into streaming response instead of parsing to standard tool_calls JSONfixed in v0.24.0
- day 56vLLM #41459fixfix(frontend): Add multimodal placeholders to Gemma4 tool message templatein v0.22.0
- day 56vLLM #43850fix[Rust Frontend] Reduce Gemma4 tool parser args scan complexityin v0.22.0
- day 43vLLM #42696bug[Bug]: Gemma4 tool parser is broken in the streaming mode (for OpenCode)fixed in v0.24.0
- day 42vLLM #42128fix[Bugfix] Fix Gemma4ToolParser streaming float corruptionin v0.21.0
- day 40vLLM #42403bug[Bug]: (Gemma 4) xgrammar bitmask lets <end_of_turn> escape during structured outputs, terminating generation mid-JSONfixed in v0.27.0
- day 40SGLang #25073bug[Bug] Gemma4 tool-call parser sets streaming `tool_calls[*].index` to the tool's position in the request, not the call indexfixed in v0.5.15.post1
- day 40vLLM #42189fix[Bugfix] [Frontend] Responses API, fix merging of messagesin v0.21.0
- day 39vLLM #42188fix[Bugfix] Gemma 4 chat template crash with missing tool name and tool idin v0.21.0
- day 36vLLM #41991fix[Bugfix][Gemma4] Fix infinite loop and array boundary issues in tool parserin v0.20.2
- day 35vLLM #41967bug[Bug]: Gemma4 + MTP speculative decoding drops first tool-call arguments in streaming multi-tool auto-tool-choicefixed in v0.24.0
- day 35SGLang #23952featurefeat(reasoning): auto-detect reasoning/tool-call parser from chat templatein v0.5.12
- day 33SGLang v0.5.11supportmodel supportin v0.5.11
- day 30vLLM #39570fix[Fix] Sync gemma4 chat template from hfin v0.20.1
- day 29vLLM #41452bug[Bug]: Gemma4-31B-it deployed on vLLM cannot process images in tool messagefixed in v0.22.0
- day 28TensorRT-LLM #13248support[None][feat] AutoDeploy: add Gemma 4 reasoning and tool-call parsersin v1.3.0rc14
- day 27XGrammar v0.1.34supporttool-calling supportin v0.1.34
- day 24vLLM #40911bug[Bug]: Tool call leaks into contentfixed in v0.24.0
- day 18vLLM #39892fix[Bugfix][Responses API] Fix streaming tool calls on /v1/responsesin v0.20.0
- day 16vLLM #39352fix[Frontend] Preserve structured output special tokens in offline LLM.chatin v0.20.0
- day 15TensorRT-LLM v1.3.0rc12supportmodel supportin v1.3.0rc12
- day 13vLLM #39948bug[Bug]: Gemma4 Auto Tool Parsing via Responses APIfixed in v0.19.1
- day 12vLLM #39679fix[Bugfix] Fix Gemma4 tool parser converting bare `null` to string `"null"`in v0.19.1
- day 11vLLM #39652bugGemma4ToolParser.__init__ missing 'tools' parameter causes TypeErrorfixed in v0.19.1
- day 11XGrammar #588supportfeat: add builtin structural tag support for Gemma 4 modelsin v0.1.34
- day 8vLLM #39468bug[Bug]: vllm 0.19.0, gemma4, The format of the tool call returned by vllm is incorrect.open
- day 7vLLM #39380bugGemma4 streaming tool call corrupts decimal numbers (X.Y → X.0Y)fixed in v0.19.1
- day 6vLLM #38909fix[Bugfix][Frontend] Fix Gemma4 streaming HTML duplication after tool callsin v0.19.1
- day 6vLLM #39027fix[Tool] `adjust_request` to reasoning parser, and Gemma4 fixesin v0.19.1
- day 6vLLM #39114fix[Bugfix] Fix Gemma4 streaming tool call corruption for split boolean/number values in v0.19.1
- day 5SGLang #21952support[New Model] Gemma 4in v0.5.10.post1
- day 4vLLM #39069bug[Bug]: gemma4_utils._parse_tool_arguments truncates string values containing internal quotesfixed in v0.24.0
- day 4vLLM #39072bug[Bug]: Gemma4 on vLLM + PI coding agent: Validation failed for tool "edit": - path: must have required property 'path'open
- day 4vLLM #39089bug[Bug]: gemma4 tool-call-parser corrupts boolean values in tool call arguments during streaming modefixed in v0.19.1
- day 4vLLM #39130bug[Bug]: `--reasoning-parser gemma4` silently disables structured output (xgrammar) when `enable_thinking=false`fixed in v0.24.0
- day 3vLLM #39043bug[Bug]: Vllm + Gemma 4 + claude code: tool calling problemsfixed in v0.25.0
- day 3vLLM #38992fix[Bugfix] Fix invalid JSON in Gemma 4 streaming tool calls by stripping partial delimitersin v0.19.1
- day 1vLLM #38910bug[Bug]: Gemma4 tool parser duplicates HTML tag prefixes in streamed tool argumentsfixed in v0.19.1
- day 1vLLM #38946bug[Bug]: Gemma 4 tool usage produces invalid json due to bug in streamingfixed in v0.19.1
- day 1vLLM v0.19.0supporttool-calling supportin v0.19.0
- day 0vLLM #38837bug[Bug]: Gemma4ToolParser.__init__() missing `tools` parameter — 400 error on tool callsfixed in v0.19.0
- day 0vLLM #38826supportfeat(models): implement Google Gemma 4 architecture support (MoE, Multimodal, Reasoning, Tool-Use)in v0.19.0
- day 0vLLM #38847fix[Bugfix]: Fix Gemma4ToolParser.__init__() missing `tools` parameterin v0.19.0
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now0
Opened / fixed0 / 0fixed: in a release
First server supportday −27TensorRT-LLM
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (2)
- day −27TensorRT-LLM v1.3.0rc27supporttool-calling supportin v1.3.0rc27
- day −36TensorRT-LLM #18804support[TRTLLM-16163][feat] Enable Nemotron 3.5 Super VL serving pathsin v1.3.0rc26
13open tool-calling issues upstreamvLLM 10 · SGLang 2 · TensorRT-LLM 1 · XGrammar 0 · oldest open 204 days
- 0113 tool-calling bugs are open upstream for Gemma 4 26B A4B (vLLM 10, SGLang 2, TensorRT-LLM 1), the oldest for 204 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix28 daysreport to release
Open now13oldest 204 days
Opened / fixed33 / 20fixed: in a release
First server supportday 0vLLM; TensorRT-LLM day 14; XGrammar day 26; SGLang day 32
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (70)
- day 207vLLM #63768fix[Rust Frontend] Accept parenthesized and name-only Gemma 4 tool calls like Pythonmerged, not released
- day 206TensorRT-LLM #20995bug[Bug]: Gemma4 tool parser raises RecursionError on arguments nested about 1000 levels deepopen
- day 203vLLM #63059fix[Rust Frontend] Accept Gemma 4 tool calls without the call prefix like Pythonmerged, not released
- day 201vLLM #62741feature[Rust Frontend] Support --exclude-tools-when-tool-choice-nonemerged, not released
- day 194vLLM #61494fix[Bugfix][Parser] Stream Gemma4 string tool arguments incrementallyin v0.32.0
- day 188vLLM #59627fix[Bugfix][Structured Output] Feed the token that implicitly ends reasoning to the grammarin v0.32.0
- day 181vLLM #59005support[Rust Frontend] Add `hf` parser for response templatesin v0.31.0
- day 175vLLM #58647bug[RFC]: Hardening /v1/messages (Anthropic API) for Claude Code across model familiesopen
- day 174XGrammar #923supportfeat: support Gemma 4 tool calling with the <|"|> string delimitermerged, not released
- day 165SGLang #39640bug[Bug] SGLANG_FORWARD_UNKNOWN_TOOLS is not honoured by four tool-call detectors on the one-shot pathopen
- day 165vLLM #53444fix[Bugfix] Handle bare and malformed tool call openers in Gemma4 parserin v0.30.0
- day 159vLLM #55417fix[Rust Frontend] Correctly parse whitespace framing in model outputin v0.29.0
- day 152vLLM #54881bug[Bug]: Gemma 4 skips reasoning and repeats JSON when json_schema is enabledfixed in v0.29.0
- day 147vLLM #54256bug[Bug]: Gemma4 parser drops tool calls when transitioning out of reasoning channel (<channel|>call:...)open
- day 146TensorRT-LLM #14290feature[TRTLLM-12758][feat] honor named function in tool_choice for non-harmony modelsin v1.3.0rc25
- day 144vLLM #53657fix[Bugfix] Handle parenthesized Gemma4 tool callsin v0.28.0
- day 143vLLM #53642bug[Bug]: Gemma4 parser swallows the rest of the turn when the model emits call:name(...) instead of call:name{...}fixed in v0.28.0
- day 142vLLM #53431bug[Bug]: Gemma4 tool parser silently drops the bare `<|tool_call>:name{...}` opener — no tool call, no content, on both streaming and non-streamingfixed in v0.30.0
- day 138SGLang #35564bug[Bug] 8 tool-call parsers give wrong or missing calls in token-by-token streamingfixed day 208, not released
- day 125vLLM #51284bug[Bug]: gemma4 parser keeps literal quote characters when the model writes array/object string values in plain JSON or Python syntax instead of <|"|>open
- day 123vLLM #50938bug[Bug]: Text emitted before the reasoning start tag escapes both the gemma4 parser and structured output, breaking json_schema responsesopen
- day 122vLLM #50927bug[Bug]: stack overflow causing system crash: Rust unbounded recursion in the Gemma4 unified parserfixed in v0.29.0
- day 118vLLM #50477bug[Bug]: gemma4 parser silently ignores named forced tool_choice on 0.26.0 (worked on 0.21.0)open
- day 115vLLM #49745other[Refactor] Remove dead code in multiple files (#49745)in v0.27.0
- day 110vLLM #49412bug[Bug]: Streaming vs non-streaming content whitespace mismatch at tool-call boundaries (ParserEngine: qwen3/nemotron_v3/seed_oss/glm47_moe/gemma4)open
- day 108vLLM #49254bug[Bug]: nvidia/Gemma-4-26B-A4B-NVFP4 model tool parser issueopen
- day 98vLLM #48217bug[Bug]: Gemma4 parser classifies plain output without channel markers differently in streaming and non-streaming modefixed in v0.26.0
- day 95vLLM #47906bug[Bug]: Gemma4 engine-based tool parser returns all bare-value arguments as strings (scilent behavior change)open
- day 91vLLM #47289fix[Rust Frontend] [Parser]Recover buffered text from incomplete tool calls at EOSin v0.25.0
- day 90vLLM #47401bug[Bug]: ][Docker] vllm/vllm-openai-cpu:v0.24.0 image missing tool_chat_template_gemma4.jinja file in examplesfixed in v0.25.0
- day 83vLLM #46602other[Rust Frontend] Migrate gemma4 to unified parserin v0.24.0
- day 76vLLM #46091fix[Bugfix] [Parser] Fix empty tool block silently dropping subsequent contentin v0.24.0
- day 75SGLang #28475bugResponses API: gemma4 tool call leaks as text under tool_choice="auto" (parsed only with "required")fixed in v0.5.18
- day 74vLLM #45553fix[Bugfix][Gemma4] Fix offline parser truncation, adjust_request token leak, and chat template syncin v0.24.0
- day 73vLLM #45588fix[Frontend] Replace legacy Gemma4 parsers with engine-based implementationin v0.24.0
- day 62vLLM #44522bugGemma-4 tool call parser leaks raw tokens (<|" and "|>) into streaming response instead of parsing to standard tool_calls JSONfixed in v0.24.0
- day 55vLLM #41459fixfix(frontend): Add multimodal placeholders to Gemma4 tool message templatein v0.22.0
- day 55vLLM #43850fix[Rust Frontend] Reduce Gemma4 tool parser args scan complexityin v0.22.0
- day 47vLLM #43244bug[Bug]: `previous_response_id` drops function_call/function_call_output from stored context in Responses APIopen
- day 42vLLM #42696bug[Bug]: Gemma4 tool parser is broken in the streaming mode (for OpenCode)fixed in v0.24.0
- day 41vLLM #42128fix[Bugfix] Fix Gemma4ToolParser streaming float corruptionin v0.21.0
- day 39vLLM #42403bug[Bug]: (Gemma 4) xgrammar bitmask lets <end_of_turn> escape during structured outputs, terminating generation mid-JSONfixed in v0.27.0
- day 35vLLM #42047bug[Bug] Gemma4ToolParser streams incorrect float values (e.g., 108.2 → 108.02)fixed in v0.21.0
- day 35vLLM #41991fix[Bugfix][Gemma4] Fix infinite loop and array boundary issues in tool parserin v0.20.2
- day 34SGLang #23952featurefeat(reasoning): auto-detect reasoning/tool-call parser from chat templatein v0.5.12
- day 32SGLang v0.5.11supportmodel supportin v0.5.11
- day 29vLLM #39570fix[Fix] Sync gemma4 chat template from hfin v0.20.1
- day 27TensorRT-LLM #13248support[None][feat] AutoDeploy: add Gemma 4 reasoning and tool-call parsersin v1.3.0rc14
- day 26XGrammar v0.1.34supporttool-calling supportin v0.1.34
- day 23vLLM #40911bug[Bug]: Tool call leaks into contentfixed in v0.24.0
- day 17vLLM #39892fix[Bugfix][Responses API] Fix streaming tool calls on /v1/responsesin v0.20.0
- day 15vLLM #39352fix[Frontend] Preserve structured output special tokens in offline LLM.chatin v0.20.0
- day 14TensorRT-LLM v1.3.0rc12supportmodel supportin v1.3.0rc12
- day 12vLLM #39948bug[Bug]: Gemma4 Auto Tool Parsing via Responses APIfixed in v0.19.1
- day 11vLLM #39679fix[Bugfix] Fix Gemma4 tool parser converting bare `null` to string `"null"`in v0.19.1
- day 10vLLM #39652bugGemma4ToolParser.__init__ missing 'tools' parameter causes TypeErrorfixed in v0.19.1
- day 10XGrammar #588supportfeat: add builtin structural tag support for Gemma 4 modelsin v0.1.34
- day 7vLLM #39468bug[Bug]: vllm 0.19.0, gemma4, The format of the tool call returned by vllm is incorrect.open
- day 5vLLM #38909fix[Bugfix][Frontend] Fix Gemma4 streaming HTML duplication after tool callsin v0.19.1
- day 5vLLM #39027fix[Tool] `adjust_request` to reasoning parser, and Gemma4 fixesin v0.19.1
- day 4SGLang #21952support[New Model] Gemma 4in v0.5.10.post1
- day 3vLLM #39069bug[Bug]: gemma4_utils._parse_tool_arguments truncates string values containing internal quotesfixed in v0.24.0
- day 3vLLM #39089bug[Bug]: gemma4 tool-call-parser corrupts boolean values in tool call arguments during streaming modefixed in v0.19.1
- day 3vLLM #39130bug[Bug]: `--reasoning-parser gemma4` silently disables structured output (xgrammar) when `enable_thinking=false`fixed in v0.24.0
- day 2vLLM #39043bug[Bug]: Vllm + Gemma 4 + claude code: tool calling problemsfixed in v0.25.0
- day 2vLLM #38992fix[Bugfix] Fix invalid JSON in Gemma 4 streaming tool calls by stripping partial delimitersin v0.19.1
- day 0vLLM #38910bug[Bug]: Gemma4 tool parser duplicates HTML tag prefixes in streamed tool argumentsfixed in v0.19.1
- day 0vLLM v0.19.0supporttool-calling supportin v0.19.0
- day −1vLLM #38855bug[Bug]: Gemma4 reasoning parser fails to separate reasoning_content — <|channel> tokens stripped before parsingfixed in v0.24.0
- day −1vLLM #38826supportfeat(models): implement Google Gemma 4 architecture support (MoE, Multimodal, Reasoning, Tool-Use)in v0.19.0
2open tool-calling issues upstreamvLLM 2 · SGLang 0 · TensorRT-LLM 0 · XGrammar 0 · oldest open 164 days
- 012 tool-calling bugs are open upstream for Kimi K2.6 (vLLM 2), the oldest for 164 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix52 daysreport to release
Open now2oldest 164 days
Opened / fixed6 / 4fixed: in a release
First server supportday 15SGLang
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (8)
- day 112vLLM #51660bug[Bug]: No viable structured-outputs backend for Kimi-K2.6 + EAGLE3 on v0.26.0: xgrammar bitmask desync kills the engine, outlines compile-timeout leaks threads until OOM, guidance rejects the slow tokenizeropen
- day 71vLLM #46610support[Frontend] Add Streaming Parser Engine and new Kimi k2.5/k2.6/k2.7 Parserin v0.25.0
- day 49vLLM #44906bug[Bug]: Kimi-k2.6 tool call parser bug when using tool_choicefixed in v0.25.0
- day 30vLLM #43203bug[Bug]: Not support role tool of imageopen
- day 25SGLang #25358bugBug: KimiK2Detector fails to parse tool calls with callNNNNN IDs in multi-turn conversationsfixed in v0.5.16
- day 25vLLM #42747bug[Bug]: Chat Completions streaming invokes tool parser despite `tool_choice="none"`fixed in v0.22.1
- day 24SGLang #25218bug[Bug] Kimi-K2.6 generates wrong tool_call from 2nd round onwardfixed in v0.5.16
- day 15SGLang v0.5.11supportmodel supportin v0.5.11
8open tool-calling issues upstreamvLLM 5 · SGLang 0 · TensorRT-LLM 2 · XGrammar 1 · oldest open 108 days
- 018 tool-calling bugs are open upstream for Qwen3 235B A22B Instruct 2507 (vLLM 5, TensorRT-LLM 2, XGrammar 1), the oldest for 108 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix32 daysreport to release
Open now8oldest 108 days
Opened / fixed16 / 8fixed: in a release
First server supportday −98vLLM; SGLang day −85; TensorRT-LLM day −69
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (57)
- day 464vLLM #64326bug[Bug]: DelegatingParser.parse ignores model_output_token_ids, so markers written as plain text become tool calls or end reasoning (non-streaming)open
- day 463vLLM #64012bug[Bug]: v0.19.0 qwen3_xml corrupts separate tool calls across chunk boundaries (legacy/backport)open
- day 461TensorRT-LLM #20954bug[Bug]: Qwen3 tool parser fails or drops content on plain answers that open with [ or {open
- day 453vLLM #62027fix[Perf][Parser] Stream Qwen3 tool argument values incrementally (12 KB call: 2.1 s to 11 ms)in v0.32.0
- day 450vLLM #59928fix[Bugfix][Parser] Recover genuine tool calls after unclosed Qwen3 <function= opener in prosein v0.32.0
- day 444vLLM #59627fix[Bugfix][Structured Output] Feed the token that implicitly ends reasoning to the grammarin v0.32.0
- day 443XGrammar #940feature[Python] Add opt-in token_markers to get_model_structural_tagmerged, not released
- day 443vLLM #59563fix[Rust Frontend] Backtrack from safe text at a complete markerin v0.32.0
- day 437vLLM #59005support[Rust Frontend] Add `hf` parser for response templatesin v0.31.0
- day 436vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 435XGrammar #939bugBuiltin structural tags admit control markers assembled from ordinary sub-tokensfixed day 443, not released
- day 428vLLM #57340fix[Rust Frontend] Build full-output grammars from initialized reasoning parsersin v0.31.0
- day 422vLLM #56635fix[Parser] Fix: correct parser frontend handling of reasoning end and boundary tokensin v0.30.0
- day 420XGrammar #892featurefeat: add a parallel_tool_calls param for builtin structural tagsin v0.2.7
- day 419vLLM #56658bug[Bug]: Qwen parser discards prose and final answers after a literal <tool_call> markeropen
- day 414vLLM #55223fix[Perf] Eliminate full-history reasoning scans for structured outputsin v0.29.0
- day 413vLLM #55633bug[Bug]: legacy qwen3_xml streaming parser emits whitespace-only content before the first tool callfixed in v0.29.0
- day 402TensorRT-LLM #14290feature[TRTLLM-12758][feat] honor named function in tool_choice for non-harmony modelsin v1.3.0rc25
- day 401TensorRT-LLM #17575fix[#17574][fix] Complete zero-argument tool calls in the streaming tool parserin v1.3.0rc25
- day 390TensorRT-LLM #17740bug[Bug]: Streaming tool parser drops all response content emitted after a completed tool callopen
- day 387TensorRT-LLM #17574bug[Bug]: Zero-argument tool calls never complete in the streaming tool parserfixed in v1.3.0rc25
- day 372TensorRT-LLM #16866fix[https://nvbugs/6240584][fix] Qwen3ToolParser: bare-JSON fallback for reasoning-preceded tool callsin v1.3.0rc23
- day 366vLLM #49412bug[Bug]: Streaming vs non-streaming content whitespace mismatch at tool-call boundaries (ParserEngine: qwen3/nemotron_v3/seed_oss/glm47_moe/gemma4)open
- day 361vLLM #48846fix[Bugfix][Tool Parser] Preserve whitespace in parameter values (MiniMax M2, Qwen3, MiniCPM5 XML)in v0.26.0
- day 359vLLM #48753bug[Bug]: Qwen3/Qwen3-Coder tool parser strips meaningful whitespace from string parameter values (breaks exact-match edit tools)open
- day 344vLLM #47137bug[Bug]: Streaming vs non-streaming tool-parser divergence on a truncated `<tool_call>` opener (ParserEngine / Qwen3)fixed in v0.29.0
- day 337vLLM #46351fixfix: stream Qwen3 tool call string argumentsin v0.24.0
- day 332vLLM #45708other[Tests] Add Qwen3 streaming parser delta boundary casesin v0.24.0
- day 332vLLM #46047fix[Bugfix] [Parser] Fix Qwen3 latent bug in partial params dropping values containing `<`in v0.24.0
- day 332vLLM #46091fix[Bugfix] [Parser] Fix empty tool block silently dropping subsequent contentin v0.24.0
- day 330SGLang #28424bug[Bug] PP is not usable with xgrammar-constrained requestsfixed in v0.5.15.post1
- day 329vLLM #45413fix[Frontend] Add Streaming Parser Engine and new Qwen3 Parserin v0.23.0
- day 326vLLM #45003feature[Frontend] Support strict mode for tool callingin v0.23.0
- day 311vLLM #43243fixfix: parse Qwen3 XML JSON arguments firstin v0.22.0
- day 310vLLM #43582fix[Rust Frontend] Add reasoning/tool parser & renderer roundtrip testsin v0.22.0
- day 304vLLM #38973fix[ToolParser][Bugfix] Re-land: Fix anyOf/oneOf/$ref type resolution in Qwen3CoderToolParser (#37831)in v0.22.0
- day 303SGLang #24251fix[RL][TITO] Preserve whitespace in reasoning parser outputsin v0.5.12.post1
- day 290SGLang #23952featurefeat(reasoning): auto-detect reasoning/tool-call parser from chat templatein v0.5.12
- day 290SGLang #23953featurefeat(constrained): two-phase reasoning grammar + --enable-strict-thinkingin v0.5.12
- day 284XGrammar #610fixfix: align builtin structural tags with chat templates for reasoning and tool callsin v0.2.0
- day 270vLLM #40090fix[Bugfix] Fix empty delta detection in Qwen3XMLToolParser streaming (#40090)in v0.19.1
- day 261vLLM #38848fix[Bugfix] Fix Qwen3 tool parser for Responses API toolsin v0.19.1
- day 253vLLM #38189fix[Tool Parser][2/3] Use self.tools instead of request.tools in tool parsers (#38189)in v0.19.0
- day 238vLLM #27599other[CI/Build] Add common tool call parser test suitein v0.18.0
- day 226XGrammar #530supportfeat: provides builtin structural tag formats for several models.in v0.1.32
- day 100TensorRT-LLM #8216support[TRTLLM-8214][feat] Support Qwen3 tool parserin v1.2.0rc2
- day 68SGLang #10550featureUse jsonschema to constrain required or specific tool choicein v0.5.3
- day 64vLLM #25203other[Test]: Hermes tool parser stream output error in Qwen3 casein v0.11.0
- day 4vLLM #21565bug[Bug]: Qwen3 failed to get function with stream and named function calling when thinking is disabledfixed in v0.10.1rc1
- day 1SGLang #8260supportPreliminary Support for Qwen3XMLDetectorin v0.4.10
- day −30SGLang #7394fix[BugFix]fix qwen25 invoke function call streaming responses with curly braces as the starting indicatorin v0.4.8
- day −48vLLM #19056bug[Bug]: Hermes tool parser stream output error in Qwen3 casefixed in v0.11.0
- day −55SGLang #6653bug[Bug] Qwen3 Structured Output won't finish until hit the `max_tokens` limitfixed in v0.4.10
- day −69TensorRT-LLM v0.20.0rc2supportmodel supportin v0.20.0rc2
- day −77vLLM #17655bug[Bug]: Qwen3 streaming output failed to return function call results when using "qwen3" reasoning parser.fixed in v0.9.1rc1
- day −85SGLang v0.4.6supportmodel supportin v0.4.6
- day −98vLLM v0.8.4supportmodel supportin v0.8.4
13open tool-calling issues upstreamvLLM 9 · SGLang 1 · TensorRT-LLM 1 · XGrammar 2 · oldest open 155 days
- 0113 tool-calling bugs are open upstream for Qwen3.6 35B A3B (vLLM 9, SGLang 1, TensorRT-LLM 1, XGrammar 2), the oldest for 155 days.
- 02vLLM v0.32.0 fixed #40875 ([Bug]: ngram speculative decoding default prompt_lookup_min=2 causes tool-call output corruption on Qwen3-class models with structured output (config-only fix: prompt_lookup_min=8)) 13 days ago: providers on an older version still have the bug.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix67 daysreport to release
Open now13oldest 155 days
Opened / fixed19 / 6fixed: in a release
First server supportday 2XGrammar; SGLang day 8; TensorRT-LLM day 79
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (30)
- day 175SGLang #45235bug[Bug] qwen3_coder: one `strict: true` tool under tool_choice="auto" makes a non-strict zero-argument tool get invented argumentsopen
- day 151vLLM #58647bug[RFC]: Hardening /v1/messages (Anthropic API) for Claude Code across model familiesopen
- day 148vLLM #58105bug[Bug]: Qwen streaming loses tool calls when text grammar permits ordinary-token delimitersopen
- day 141vLLM #48200fix[Refactor] StructuredOutputManager x Speculative Decoding Refactorin v0.30.0
- day 121vLLM #48922fix[Bugfix] Guard tool call argument JSON parsing in chat message postprocessingin v0.29.0
- day 116TensorRT-LLM #18084bugMODEL_TYPE_TO_TOOL_PARSER maps qwen3_5/qwen3_5_moe to a parser whose format the template does not emitopen
- day 113TensorRT-LLM #17917bugA tool parser that matches its marker but extracts no calls returns 200 with no diagnosticfixed in v1.3.0rc27
- day 99vLLM #50989bug[Bug]: Qwen 3.* models go into doom loop if you use strict mode with tools without arguments.open
- day 99XGrammar #802bug[Bug] XML tool calling loops forever on no-argument toolsopen
- day 88vLLM #49724bug[Bug]: Responses custom tools are coerced to function_call on non-Harmony Qwen routesopen
- day 87SGLang #32255bug[Bug] OpenAI and Anthropic protocols are not working in sglangfixed in v0.5.16
- day 87vLLM #44993fix[Bugfix][Structured Output][Spec Decode] Advance grammar across reasoning boundaryin v0.26.0
- day 84vLLM #49205bug[Bug]: StreamingParserEngine leaks a bare tag fragment (e.g. '<tool_') into content when a PreLexedTerminal supersedes a buffered text-token prefix of the same tagopen
- day 84vLLM #49216bug[Bug]: Tools streaming internal error when `VLLM_ENFORCE_STRICT_TOOL_CALLING=false` and `tool_choice=required`open
- day 80vLLM #48874bug[Bug]: Anthropic /v1/messages renders system-role messages inside `messages` positionally into the chat template (breaks Claude Code >=2.1.2xx tool calling)open
- day 79TensorRT-LLM v1.3.0rc21supportmodel supportin v1.3.0rc21
- day 60SGLang #29441bug[Bug] Empty content chunk in SSE stream before tool call chunks breaks AI SDK providersfixed in v0.5.19
- day 46vLLM #45003feature[Frontend] Support strict mode for tool callingin v0.23.0
- day 39SGLang #27336bug[Bug] Qwen3.6 tool call with tool_choice="required" causes </parameter to appear in tool call responsefixed in v0.5.17
- day 39XGrammar #650bug[Bug] XML parameter values can drift past the end-marker prefix under constrained decoding (e.g. `</parameter1>`)open
- day 39vLLM #44676bug[Bug]: thinking_token_budget forces reasoning_end tokens into the middle of tool-call arguments — ThinkingBudgetStateHolder doesn't treat <tool_call> as implicit reasoning end (Qwen3.5+)open
- day 33SGLang #26790bugBug: Qwen tool call parser (qwen25/qwen) returns empty tool_calls list in OpenAI API responsefixed in v0.5.17
- day 32vLLM #44012bug[Bug]: Responses API + `text.format` json_schema is not grammar-constrained when a reasoning parser is enabled — output escapes into the unconstrained `reasoning` channelopen
- day 8SGLang v0.5.11supportmodel supportin v0.5.11
- day 7SGLang #21722supportfeat: use structural tags to enable strict tool calling and reasoning for more modelsin v0.5.11
- day 7vLLM #40894supportfeat: update xgrammar==0.2.0 to use structural tags for strict tool calling + reasoning for more modelsin v0.20.2
- day 2XGrammar v0.1.34supporttool-calling supportin v0.1.34
- day 2XGrammar #609fixrefactor: unify reasoning parameter and rename model keysin v0.2.0
- day 1XGrammar #603supportfeat: support Qwen3.6's tool-calling format.in v0.1.34
- day −2vLLM #40875bug[Bug]: ngram speculative decoding default prompt_lookup_min=2 causes tool-call output corruption on Qwen3-class models with structured output (config-only fix: prompt_lookup_min=8)fixed in v0.32.0
11open tool-calling issues upstreamvLLM 5 · SGLang 4 · TensorRT-LLM 2 · XGrammar 0 · oldest open 235 days
- 0111 tool-calling bugs are open upstream for DeepSeek V3.2 (vLLM 5, SGLang 4, TensorRT-LLM 2), the oldest for 235 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix21 daysreport to release
Open now11oldest 235 days
Opened / fixed30 / 19fixed: in a release
First server supportday −56SGLang; TensorRT-LLM day −24; vLLM day −13
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (101)
- day 333TensorRT-LLM #21072fix[None][fix] Responses: detokenize for the parsers as chat completions doesmerged, not released
- day 332vLLM #58780fix[Bugfix][Parser] Stream DeepSeek V4/V3.2 string tool arguments incrementallymerged, not released
- day 330TensorRT-LLM #20873fix[None][fix] DeepSeek DSML tool parsers: the layout newlines around tool calls are not contentin v1.4.0rc2
- day 330TensorRT-LLM #21025fix[None][fix] DeepSeek DSML tool parsers: keep string parameter values as writtenin v1.4.0rc2
- day 330TensorRT-LLM #21051fix[None][fix] DeepSeek renderers: tools join the leading system message; V4.1 renders developer as systemin v1.4.0rc2
- day 329TensorRT-LLM #20989bug[Performance]: DeepSeek streaming tool parsers rescan the whole tool call on every increment (30 s of CPU for one 64k-character call)open
- day 329TensorRT-LLM #21014bug[Bug]: DeepSeek V4.1 / V3.2 tool parsers cut a parameter value at a literal DSML closing tag and drop or leak the rest of the callopen
- day 326vLLM #63314fix[Rust Frontend] Render empty tool call arguments as an empty object like Pythonmerged, not released
- day 324vLLM #62741feature[Rust Frontend] Support --exclude-tools-when-tool-choice-nonemerged, not released
- day 321SGLang #44941bugDeepSeek V3.2/V4/V4.1 streaming tool-call parser: pre-existing bugs found while fuzzingopen
- day 303vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 299SGLang #41317bug[Bug] DeepSeek V3.2/V4 DSML detector strips whitespace (e.g. trailing newline) from string="true" parameter valuesopen
- day 298vLLM #58647bug[RFC]: Hardening /v1/messages (Anthropic API) for Claude Code across model familiesopen
- day 295SGLang #39632fixfix(function_call): buffer complete DeepSeek DSML invokesin v0.5.21
- day 293vLLM #57826bug[Bug]: minimax_m2 / glm47_moe / deepseek_v4 arg converters drop the last parameter when the model omits the closing tag (same class as #57699)open
- day 291SGLang #40236bug[Bug] DeepSeek V4/V3.2: DSML tool calls returned as content with no tool_calls (bare invoke, unterminated section, malformed sibling)open
- day 288vLLM #56200fix[Refactor] Derive is_reasoning_end from the engine grammarin v0.30.0
- day 283SGLang #38924bug[Bug] DeepSeek V4/V3.2 DSML tool-call parser occasionally wraps arguments in a spurious "arguments"/"input" keyopen
- day 282vLLM #55954fix[Bugfix] Parse DSML tool calls when the model omits the tool_calls wrapperin v0.29.0
- day 275vLLM #54838fix[Bugfix] Implicitly close DeepSeek DSML parametersin v0.29.0
- day 268TensorRT-LLM #17575fix[#17574][fix] Complete zero-argument tool calls in the streaming tool parserin v1.3.0rc25
- day 261SGLang #35563bug[Bug] deepseekv32/deepseekv4 tool-call parsers drop calls when the streamed output has two chunksfixed in v0.5.18
- day 261TensorRT-LLM #17903fix[#17580][fix] Stream the text that precedes a tool call in DeepSeek parsersin v1.3.0rc25
- day 260TensorRT-LLM #17573fix[#17572][fix] Emit the withheld buffer in DeepSeek streaming tool parsersin v1.3.0rc25
- day 254TensorRT-LLM #17572bug[Bug]: DeepSeek streaming tool parsers drop text they withheld while waiting on a tool-call tokenfixed in v1.3.0rc25
- day 254TensorRT-LLM #17580bug[Bug]: DeepSeek streaming tool parsers drop text that precedes a tool call in the same deltafixed in v1.3.0rc25
- day 253vLLM #51144feature[Rust Frontend] Support dynamic tools from developer messagesin v0.27.1
- day 252vLLM #51727fix[Bugfix] Fix DeepSeek V4/3.2 tokenizer vocab size overcount crashing guided decodingin v0.27.1
- day 221SGLang #29883fix[BUG] fix strip streaming empty-string suffix from DSV4 tool argumentsin v0.5.15
- day 220SGLang #30623other[Refactor] Share chat encoding dispatch between serving and offline toolsin v0.5.15
- day 219vLLM #47986bug[Bug]: DeepSeek V4/V32 parser can unwrap arguments using the wrong tool's schema when parallel tool calls share identical raw argument textopen
- day 211vLLM #47179bug[Rust Frontend] Tool-call argument parsing diverges from the Python parsers (coercion, whitespace, unions)open
- day 204SGLang #28449supportExpand parser auto detection coveragein v0.5.14
- day 199vLLM #43984fix[Bugfix][Tool Parser] Handle non-finite numbers in coerce_to_schema_typein v0.24.0
- day 193vLLM #45003feature[Frontend] Support strict mode for tool callingin v0.23.0
- day 179SGLang #25676supportUpgrade xgrammar to 0.2.1in v0.5.13
- day 178vLLM #42879fix[Bugfix] Stream DeepSeek DSML tool-call argument deltas incrementallyin v0.22.0
- day 178TensorRT-LLM #14611fix[None][feat] Support tool-calling in KvCacheAwareRouter for disagg serving (feat/deepseek_v4)in v1.3.0rc17
- day 171vLLM #43255other[CI] Add composed-schema regression tests for DeepSeek V3.2/V4 parsersin v0.22.0
- day 170vLLM #43019fix[Bugfix] Use shared coerce_to_schema_type in DeepSeekV32 tool parserin v0.22.0
- day 167XGrammar #638fixfix(structural_tag): align deepseek_v3_2/v4 invoke separator with chat templatein v0.2.1
- day 165SGLang #25233fix[Fix] DeepSeek-V3.2: build structural tag locally to encode both wrapper and invoke layersin v0.5.12
- day 164XGrammar #637bugdeepseek_v3_2 structural tag: double-newline between invoke blocks blocks parallel tool calls at greedy decodingfixed in v0.2.1
- day 156vLLM #41801fix[Bugfix] DeepSeekV32/v4: respect string='true|false' attribute andunwrap arguments/input wrapperin v0.20.2
- day 151XGrammar #610fixfix: align builtin structural tags with chat templates for reasoning and tool callsin v0.2.0
- day 149vLLM #41198fix[Bugfix] DSV32/V4 add missing type conversion for non-streaming tool callsin v0.20.1
- day 148vLLM #41132bug[Bug]: DeepSeek V3.2 & V4 incorrect structured output when thinking enabledfixed in v0.20.1
- day 148SGLang #23915fixFix dsv4 self-closing invoke tags + accept reasoning_effort=maxin v0.5.11
- day 146vLLM #40806fix[Bugfix] Fix the DSML token leakage in DSV4/3.2in v0.20.0
- day 131SGLang #21593fixFix tool call constrained decoding and parsing for models with native formatsin v0.5.11
- day 121vLLM #38172other[Misc] Add 20 regression tests for 11 tool parser bug fixesin v0.19.0
- day 120vLLM #38189fix[Tool Parser][2/3] Use self.tools instead of request.tools in tool parsers (#38189)in v0.19.0
- day 119vLLM #33703fix[Bugfix] Support multi-type params parsing for DeepSeek v3.2in v0.18.1
- day 112SGLang #21176bug[Bug] DeepSeek-V3.2 tool-call parsing exception or empty output in Disaggregated (PD) mode with SGLang 0.5.9fixed in v0.5.12.post1
- day 108vLLM #36056fix[Bugfix] Fix Deepseekv32 tool parser when stream interval > 1in v0.18.0
- day 108TensorRT-LLM #12279other[https://nvbugs/5937478][test] Add RCCA test for DeepSeek-V3.2 multi-turn tool_call encodingin v1.3.0rc9
- day 102TensorRT-LLM #12160fix[https://nvbugs/5937478][fix] Fix DS v32 tool calling type and parse errorin v1.3.0rc8
- day 99vLLM #36654bug[Bug]: Frequent Tool Call Parsing Failures with DeepSeek-V3.2open
- day 99TensorRT-LLM #11935fix[https://nvbugs/5937478][fix] Fix DS v32 tool calling type and parse errorin v1.3.0rc7
- day 92SGLang #16091feature[Tool Call] Stream DeepSeek-V3.2 function call parameters in JSON format.in v0.5.10rc0
- day 92SGLang #18174fix[Bugfix] Catch errors when DeepSeek-V3.2 generates malformed JSONin v0.5.10rc0
- day 88SGLang #19370fixFix parallel tool call test for speculative decoding variantsin v0.5.10rc0
- day 78XGrammar #529supportfeat: support MinimaxXmlFormat and DeepSeekXmlFormat.in v0.1.32
- day 73SGLang #18302otheradd tool_choice=auto nightly test casein v0.5.9
- day 69vLLM #33855other[Perf] Simplify DeepseekV32 tokenizer, ensure fast detokenization usedin v0.16.0
- day 67vLLM #33964fix[Bugfix] Fix the issue where tool calling does not work when using fast detokenization with dsv32in v0.16.0
- day 59SGLang #17942bug[Bug] DeepSeek-V3.2 tool calls incorrectly output to content field instead of tool_calls field withfixed in v0.5.9
- day 58vLLM #33215bug[Bug]: DeepSeek V3.2 `tool_choice==required` in thinking mode gives internal server error.fixed in v0.16.0
- day 52SGLang #17551bug[Bug] DeepSeek V3.2: All output marked as reasoning_content after tool call resultfixed in v0.5.9
- day 52SGLang #17561bug[Bug] DeepSeek-V3.2 Places Tool Call Content in "content" Field Instead of "tool_calls" When Invoking Tools, and finish_reason is "stop" Rather Than "tool_calls"fixed in v0.5.9
- day 52SGLang #17593bug[Bug] DeepSeek-V3.2 tool calls incorrectly output to content field instead of tool_calls fieldfixed in v0.5.9
- day 46SGLang #17227bug[Bug] Wrong template for DeepSeek 3.2?fixed in v0.5.8
- day 43vLLM #32240other[Refactor] [6/N] to simplify the vLLM openai chat_completion serving architecture (#32240)in v0.14.0
- day 40SGLang #16698fixfix(function_call): group batch decode by options instead of fallbackin v0.5.8
- day 25SGLang #14750fix[Tool Call][DSV32] Streamline function call parametersin v0.5.7
- day 22TensorRT-LLM #10126support[TRTLLM-9677][feat] Support DeepSeek-V3.2 tool parserin v1.2.0rc6
- day 18TensorRT-LLM #9814support[TRTLLM-9654][feat] Support DeepSeek-V32 chat templatein v1.1.0
- day 17SGLang #15278featurefeat: DeepSeek-V3.2 Streaming tool call outputin gateway-v0.3.0
- day 16SGLang #15155fixfix(function_call): fallback to decode when batch decode options differin gateway-v0.3.0
- day 15SGLang #15092fix[Bugfix][Tool Call] Add null system prompt to support tool system promptin gateway-v0.3.0
- day 14vLLM #30671fix[Bugfix] Fix missing first token in tool calls during reasoning-to-tool transitionin v0.13.0
- day 14vLLM #30675other[Refactor] [2/N] Move tool parsers into the vLLM main directory (#30675)in v0.13.0
- day 12SGLang #15042bug[Bug] deepseek v3.2 Under high concurrency, the function calling feature is experiencing issues.fixed in v0.5.9
- day 12SGLang #15050bug[Bug] GrammarMatcher error with EagleWorkerfixed in v0.5.9
- day 12SGLang #15064fixfix: dpskv32 chat history processing, default drop_thinking to truein gateway-v0.3.0
- day 9vLLM #30371fix[Bugfix] Fix the issue where DeepSeek v3.2 cannot use structured_outputin v0.13.0
- day 8SGLang #14690bug[Bug] does DeepSeek-V3.2 tool-call-parser is deepseekv31?fixed in gateway-v0.2.4
- day 8SGLang #14695bug[Bug] DeepSeek-V3.2 occasional malformed tool call output (missing ｜DSML｜ markers with --tool-call-parser deepseekv32)fixed in v0.5.8
- day 8SGLang #14711bug[Bug] deepseek v3.2 function call cannot output in a streaming mannerfixed in gateway-v0.3.0
- day 7SGLang #14598bug[Bug] Sglang 0.5.6 not supports toolcall correctfixed in gateway-v0.2.4
- day 7SGLang #14573fix[Tool Call] Fix DeepSeekV32Detector skipping functions with no params in streaming modein gateway-v0.2.4
- day 3vLLM #29848supportAdd DeepSeek-V3.2 tool parser.in v0.13.0
- day 3vLLM #30025fix[Bugfix] fixed deepseekv32 tool calling errorin v0.13.0
- day 2vLLM #29837support[Frontend] supports deepseekv32 chat templatein v0.13.0
- day 1SGLang #14249supportfeat: DeepSeek new v3.2 encodingin v0.5.6
- day −13vLLM v0.11.1supportmodel supportin v0.11.1
- day −24TensorRT-LLM v1.2.0rc2supportmodel supportin v1.2.0rc2
- day −32SGLang #12123fixFix DeepSeek chat templates to handle tool call arguments type checking (#11700)in v0.5.5
- day −46SGLang #11700bug[Bug] DeepSeekV3 series model chat_template contains redundant tojson in tool_calls, affecting Multi-Round FunctionCallfixed in v0.5.5
- day −56SGLang v0.5.3supportmodel supportin v0.5.3
- day −57SGLang #11063supportAdd DeepSeek-V3.2 Tool Call Templatein v0.5.3
1open tool-calling issue upstreamvLLM 1 · SGLang 0 · TensorRT-LLM 0 · XGrammar 0 · oldest open 15 days
- 011 tool-calling bug is open upstream for Nemotron 3.5 Lightning (vLLM 1), the oldest for 15 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now1oldest 15 days
Opened / fixed1 / 0fixed: in a release
First server supportday 11SGLang
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (2)
- day 66vLLM #62078bugNemotron-3.5 tool calls lose trailing/edge newlines under the wrapped qwen3_coder format (_trim_wrapping_newlines) — needs a compact converter like #58019open
- day 11SGLang v0.5.18supportmodel supportin v0.5.18
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now0
Opened / fixed0 / 0fixed: in a release
First server support—no support change found
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (0)
3open tool-calling issues upstreamvLLM 0 · SGLang 3 · TensorRT-LLM 0 · XGrammar 0 · oldest open 57 days
- 013 tool-calling bugs are open upstream for DeepSeek V4 Flash Vision Exp (SGLang 3), the oldest for 57 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now3oldest 57 days
Opened / fixed3 / 0fixed: in a release
First server supportday 32vLLM
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (4)
- day 32vLLM v0.30.0supportmodel supportin v0.30.0
- day 20SGLang #38924bug[Bug] DeepSeek V4/V3.2 DSML tool-call parser occasionally wraps arguments in a spurious "arguments"/"input" keyopen
- day 18SGLang #38450bug[Bug] DeepSeek-V4-Flash-Vision preview image: multi-turn tool calls come back wrapped in {"arguments": {...}}open
- day 14SGLang #38013bug[Bug]open
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now0
Opened / fixed0 / 0fixed: in a release
First server support—no support change found
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (1)
- day 9SGLang #45257support[Feat] parser: Add Beam reasoning, tool calls and response templates (5/7)merged, not released
1open tool-calling issue upstreamvLLM 1 · SGLang 0 · TensorRT-LLM 0 · XGrammar 0 · oldest open 2 days
- 011 tool-calling bug is open upstream for Muse Spark Open (vLLM 1), the oldest for 2 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now1oldest 2 days
Opened / fixed1 / 0fixed: in a release
First server support—no support change found
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (5)
- day 14vLLM #64517bug[Bug]: Muse, MiniCPM5 and Poolside V1 tool parsers JSON-decode string parameters not declared as plain "type": "string" (e.g. Optional[str])open
- day 8SGLang #44685fix[Fix] muse: Support Responses namespace tools and defer_loadingmerged, not released
- day 7TensorRT-LLM #20770support[None][feat] Muse Spark Open modeling_v2 TEP target: mid-high-concurrency optimizations for GB300in v1.4.0rc2
- day 4vLLM #62207support[Frontend] Add Muse Spark Open reasoning and tool parsersmerged, not released
- day 3SGLang #44666support[Feat] srt: Add muse spark runtime, multimodal support and tests (3/5) (#44666)in v0.5.22
1open tool-calling issue upstreamvLLM 1 · SGLang 0 · TensorRT-LLM 0 · XGrammar 0 · oldest open 35 days
- 011 tool-calling bug is open upstream for Llama 3.3 70B Instruct (vLLM 1), the oldest for 35 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix32 daysreport to release
Open now1oldest 35 days
Opened / fixed5 / 4fixed: in a release
First server supportday 433TensorRT-LLM
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (8)
- day 659vLLM #58824bug[Bug]: llama3_json streaming drops assistant content that starts with '{' but is not a tool call (non-streaming returns it)open
- day 504vLLM #40059fix[BUG]: fix HF tokenizer concurrent borrow in tool parsersin v0.20.0
- day 433TensorRT-LLM v1.3.0rc3supportmodel supportin v1.3.0rc3
- day 138SGLang #5661bug[Bug] Streaming Response Missing index in tool_calls with LLaMA-3.3-70B using sglangfixed in v0.4.7
- day 107SGLang #4700bug[Bug] Missing tool name in answerfixed in v0.4.5
- day 104vLLM #15236bug[Bug]: Major issues with guided generation / structured output in vLLM (up to and including v0.8.1); many examples provided by vllm in /examples and structured_outputs.html doc do not workfixed in v0.8.3rc1
- day 97SGLang #4320fixFix Llama3.3 tool call supportin v0.4.5
- day 88SGLang #4072bug[Bug] Tool call with Llama3 models has inconsistent behavior with OpenAIfixed in v0.4.7
1open tool-calling issue upstreamvLLM 1 · SGLang 0 · TensorRT-LLM 0 · XGrammar 0 · oldest open 101 days
- 011 tool-calling bug is open upstream for Nemotron 3 Ultra (vLLM 1), the oldest for 101 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now1oldest 101 days
Opened / fixed1 / 0fixed: in a release
First server supportday −668vLLM; SGLang day 9
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (4)
- day 48vLLM #49412bug[Bug]: Streaming vs non-streaming content whitespace mismatch at tool-call boundaries (ParserEngine: qwen3/nemotron_v3/seed_oss/glm47_moe/gemma4)open
- day 39SGLang #30533fixmore fixes for Nemotron 3 parser for tool call and force nonempty contentin v0.5.15.post1
- day 9SGLang v0.5.13supportmodel supportin v0.5.13
- day −668vLLM v0.5.4supportmodel supportin v0.5.4
1open tool-calling issue upstreamvLLM 1 · SGLang 0 · TensorRT-LLM 0 · XGrammar 0 · oldest open 90 days
- 011 tool-calling bug is open upstream for Kimi K2.7 Code (vLLM 1), the oldest for 90 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix72 daysreport to release
Open now1oldest 90 days
Opened / fixed3 / 2fixed: in a release
First server supportday 14SGLang
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (5)
- day 81vLLM #54701bug[Bug] kimi_k2 streaming tool parser intermittently emits empty tool_calls deltas under concurrent load (finish_reason=tool_calls, no name/arguments)fixed in v0.31.0
- day 51vLLM #50768bug[Bug]: Responses API Returns Raw Malformed Kimi Tool Marker as Assistant Text Instead of function_callopen
- day 18vLLM #46610support[Frontend] Add Streaming Parser Engine and new Kimi k2.5/k2.6/k2.7 Parserin v0.25.0
- day 14SGLang v0.5.14supportmodel supportin v0.5.14
- day 4vLLM #45866bug[Bug]: Kimi K2 - stream_interval 4 + kimi_k2_reasoning_parserfixed in v0.31.0
8open tool-calling issues upstreamvLLM 5 · SGLang 0 · TensorRT-LLM 2 · XGrammar 1 · oldest open 108 days
- 018 tool-calling bugs are open upstream for Qwen3 30B A3B Instruct 2507 (vLLM 5, TensorRT-LLM 2, XGrammar 1), the oldest for 108 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix36 daysreport to release
Open now8oldest 108 days
Opened / fixed22 / 14fixed: in a release
First server supportday −106vLLM; SGLang day −93; TensorRT-LLM day −77
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (62)
- day 456vLLM #64326bug[Bug]: DelegatingParser.parse ignores model_output_token_ids, so markers written as plain text become tool calls or end reasoning (non-streaming)open
- day 455vLLM #64012bug[Bug]: v0.19.0 qwen3_xml corrupts separate tool calls across chunk boundaries (legacy/backport)open
- day 453TensorRT-LLM #20954bug[Bug]: Qwen3 tool parser fails or drops content on plain answers that open with [ or {open
- day 445vLLM #62027fix[Perf][Parser] Stream Qwen3 tool argument values incrementally (12 KB call: 2.1 s to 11 ms)in v0.32.0
- day 442vLLM #59928fix[Bugfix][Parser] Recover genuine tool calls after unclosed Qwen3 <function= opener in prosein v0.32.0
- day 436vLLM #59627fix[Bugfix][Structured Output] Feed the token that implicitly ends reasoning to the grammarin v0.32.0
- day 436vLLM #60354fix[Bugfix][ToolParser] Stream every tool call when one delta completes severalin v0.32.0
- day 435XGrammar #940feature[Python] Add opt-in token_markers to get_model_structural_tagmerged, not released
- day 435vLLM #59563fix[Rust Frontend] Backtrack from safe text at a complete markerin v0.32.0
- day 429vLLM #59005support[Rust Frontend] Add `hf` parser for response templatesin v0.31.0
- day 428vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 427XGrammar #939bugBuiltin structural tags admit control markers assembled from ordinary sub-tokensfixed day 435, not released
- day 420vLLM #57340fix[Rust Frontend] Build full-output grammars from initialized reasoning parsersin v0.31.0
- day 414vLLM #56635fix[Parser] Fix: correct parser frontend handling of reasoning end and boundary tokensin v0.30.0
- day 412XGrammar #892featurefeat: add a parallel_tool_calls param for builtin structural tagsin v0.2.7
- day 411vLLM #56658bug[Bug]: Qwen parser discards prose and final answers after a literal <tool_call> markeropen
- day 406vLLM #55223fix[Perf] Eliminate full-history reasoning scans for structured outputsin v0.29.0
- day 405vLLM #55633bug[Bug]: legacy qwen3_xml streaming parser emits whitespace-only content before the first tool callfixed in v0.29.0
- day 394TensorRT-LLM #14290feature[TRTLLM-12758][feat] honor named function in tool_choice for non-harmony modelsin v1.3.0rc25
- day 393TensorRT-LLM #17575fix[#17574][fix] Complete zero-argument tool calls in the streaming tool parserin v1.3.0rc25
- day 382TensorRT-LLM #17740bug[Bug]: Streaming tool parser drops all response content emitted after a completed tool callopen
- day 379TensorRT-LLM #17574bug[Bug]: Zero-argument tool calls never complete in the streaming tool parserfixed in v1.3.0rc25
- day 364TensorRT-LLM #16866fix[https://nvbugs/6240584][fix] Qwen3ToolParser: bare-JSON fallback for reasoning-preceded tool callsin v1.3.0rc23
- day 358vLLM #49412bug[Bug]: Streaming vs non-streaming content whitespace mismatch at tool-call boundaries (ParserEngine: qwen3/nemotron_v3/seed_oss/glm47_moe/gemma4)open
- day 353vLLM #48846fix[Bugfix][Tool Parser] Preserve whitespace in parameter values (MiniMax M2, Qwen3, MiniCPM5 XML)in v0.26.0
- day 351vLLM #48753bug[Bug]: Qwen3/Qwen3-Coder tool parser strips meaningful whitespace from string parameter values (breaks exact-match edit tools)open
- day 336vLLM #47137bug[Bug]: Streaming vs non-streaming tool-parser divergence on a truncated `<tool_call>` opener (ParserEngine / Qwen3)fixed in v0.29.0
- day 329vLLM #46351fixfix: stream Qwen3 tool call string argumentsin v0.24.0
- day 324vLLM #45708other[Tests] Add Qwen3 streaming parser delta boundary casesin v0.24.0
- day 324vLLM #46047fix[Bugfix] [Parser] Fix Qwen3 latent bug in partial params dropping values containing `<`in v0.24.0
- day 324vLLM #46091fix[Bugfix] [Parser] Fix empty tool block silently dropping subsequent contentin v0.24.0
- day 321vLLM #45413fix[Frontend] Add Streaming Parser Engine and new Qwen3 Parserin v0.23.0
- day 318vLLM #45003feature[Frontend] Support strict mode for tool callingin v0.23.0
- day 303vLLM #43243fixfix: parse Qwen3 XML JSON arguments firstin v0.22.0
- day 302vLLM #43582fix[Rust Frontend] Add reasoning/tool parser & renderer roundtrip testsin v0.22.0
- day 296vLLM #38973fix[ToolParser][Bugfix] Re-land: Fix anyOf/oneOf/$ref type resolution in Qwen3CoderToolParser (#37831)in v0.22.0
- day 295SGLang #24251fix[RL][TITO] Preserve whitespace in reasoning parser outputsin v0.5.12.post1
- day 276XGrammar #610fixfix: align builtin structural tags with chat templates for reasoning and tool callsin v0.2.0
- day 262vLLM #40090fix[Bugfix] Fix empty delta detection in Qwen3XMLToolParser streaming (#40090)in v0.19.1
- day 253vLLM #38848fix[Bugfix] Fix Qwen3 tool parser for Responses API toolsin v0.19.1
- day 245vLLM #38189fix[Tool Parser][2/3] Use self.tools instead of request.tools in tool parsers (#38189)in v0.19.0
- day 230vLLM #27599other[CI/Build] Add common tool call parser test suitein v0.18.0
- day 218XGrammar #530supportfeat: provides builtin structural tag formats for several models.in v0.1.32
- day 92TensorRT-LLM #8216support[TRTLLM-8214][feat] Support Qwen3 tool parserin v1.2.0rc2
- day 60SGLang #10550featureUse jsonschema to constrain required or specific tool choicein v0.5.3
- day 56vLLM #25203other[Test]: Hermes tool parser stream output error in Qwen3 casein v0.11.0
- day 29SGLang #9696bug[Bug] parallel_tool_calls=False is not strictly enforcedfixed in gateway-v0.3.1
- day 23vLLM #23315bug[Bug]: Structured Output Not Supported in Qwen3-30B-A3B-Instruct-2507fixed in v0.10.2
- day 16SGLang #9184bug[Bug] Qwen3-30b-A3B use qwen25 tool parse, the tool_call tags occasionally show up in “content”fixed in v0.5.1
- day −4vLLM #21565bug[Bug]: Qwen3 failed to get function with stream and named function calling when thinking is disabledfixed in v0.10.1rc1
- day −7vLLM #21372bug[Bug]: Tool call argument value of type `integer` may break things when `stream=True`fixed in v0.11.1
- day −7SGLang #8260supportPreliminary Support for Qwen3XMLDetectorin v0.4.10
- day −25SGLang #7769bugQwen3-30B-A3B tool-call-parserfixed in v0.5.3
- day −38SGLang #7394fix[BugFix]fix qwen25 invoke function call streaming responses with curly braces as the starting indicatorin v0.4.8
- day −56vLLM #19051bug[Bug]: 400 response when using Qwen3 + reasoning + tool calling + tool_choice "required"fixed in v0.9.1rc1
- day −56vLLM #19056bug[Bug]: Hermes tool parser stream output error in Qwen3 casefixed in v0.11.0
- day −62vLLM #18819bug[Bug]: Broken Structured Output (Guided Decoding) with Qwen3 models when `enable_thinking=False`fixed in v0.9.2rc1
- day −63SGLang #6675bug[Bug] Broken Structured Outputs with enable_thinking=False in Qwen3fixed in v0.5.1
- day −77TensorRT-LLM v0.20.0rc2supportmodel supportin v0.20.0rc2
- day −85vLLM #17655bug[Bug]: Qwen3 streaming output failed to return function call results when using "qwen3" reasoning parser.fixed in v0.9.1rc1
- day −93SGLang v0.4.6supportmodel supportin v0.4.6
- day −106vLLM v0.8.4supportmodel supportin v0.8.4
1open tool-calling issue upstreamvLLM 1 · SGLang 0 · TensorRT-LLM 0 · XGrammar 0 · oldest open 36 days
- 011 tool-calling bug is open upstream for Qwen3.8 2.4T A95B (vLLM 1), the oldest for 36 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix2 daysreport to release
Open now1oldest 36 days
Opened / fixed2 / 1fixed: in a release
First server supportday 10SGLang
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (3)
- day 52vLLM #59834bug[Bug]: Responses API streaming regenerates output item id/call_id in response.completed (breaks strict clients)fixed in v0.31.0
- day 44vLLM #58647bug[RFC]: Hardening /v1/messages (Anthropic API) for Claude Code across model familiesopen
- day 10SGLang v0.5.18supportmodel supportin v0.5.18
1open tool-calling issue upstreamvLLM 1 · SGLang 0 · TensorRT-LLM 0 · XGrammar 0 · oldest open 154 days
- 011 tool-calling bug is open upstream for MiniMax M2.7 (vLLM 1), the oldest for 154 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now1oldest 154 days
Opened / fixed1 / 0fixed: in a release
First server supportday 42TensorRT-LLM
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (3)
- day 73vLLM #44060bugMiniMax-M2.1/M2.5/M2.7 tool-call arguments cannot safely contain raw </parameter> delimitersopen
- day 42TensorRT-LLM v1.3.0rc13supportmodel supportin v1.3.0rc13
- day 29vLLM #39861fix[Bugfix] Accept **kwargs in MiniMaxM2Parser.__init__()in v0.19.1
6open tool-calling issues upstreamvLLM 5 · SGLang 0 · TensorRT-LLM 0 · XGrammar 1 · oldest open 220 days
- 016 tool-calling bugs are open upstream for Qwen3.5 397B A17B (vLLM 5, XGrammar 1), the oldest for 220 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix9 daysreport to release
Open now6oldest 220 days
Opened / fixed13 / 7fixed: in a release
First server supportday 8SGLang; TensorRT-LLM day 8; vLLM day 19
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (28)
- day 233XGrammar #940feature[Python] Add opt-in token_markers to get_model_structural_tagmerged, not released
- day 226vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 225XGrammar #939bugBuiltin structural tags admit control markers assembled from ordinary sub-tokensfixed day 233, not released
- day 221vLLM #58647bug[RFC]: Hardening /v1/messages (Anthropic API) for Claude Code across model familiesopen
- day 210XGrammar #892featurefeat: add a parallel_tool_calls param for builtin structural tagsin v0.2.7
- day 201vLLM #55495bug[Bug]: qwen3_xml tool parser emits a truncated `arguments` string with a leaked `</parameter` tag; the Responses API then rejects every later request of the conversation with `400 Expecting value`open
- day 191TensorRT-LLM #17575fix[#17574][fix] Complete zero-argument tool calls in the streaming tool parserin v1.3.0rc25
- day 162TensorRT-LLM #16866fix[https://nvbugs/6240584][fix] Qwen3ToolParser: bare-JSON fallback for reasoning-preceded tool callsin v1.3.0rc23
- day 157vLLM #44993fix[Bugfix][Structured Output][Spec Decode] Advance grammar across reasoning boundaryin v0.26.0
- day 154vLLM #49205bug[Bug]: StreamingParserEngine leaks a bare tag fragment (e.g. '<tool_') into content when a PreLexedTerminal supersedes a buffered text-token prefix of the same tagopen
- day 120SGLang #28407bug[Bug] NEXTN (MTP) + structured output (xgrammar) crashes scheduler on Mamba-hybrid model (Qwen3.5-397B) — TypeError 'NoneType' in prepare_mamba_track_for_verify (sm_100)fixed in v0.5.18
- day 116vLLM #45003feature[Frontend] Support strict mode for tool callingin v0.23.0
- day 100vLLM #43582fix[Rust Frontend] Add reasoning/tool parser & renderer roundtrip testsin v0.22.0
- day 93vLLM #43220bug[Bug]: Streaming reasoning tokens truncated when `</think>` and `<tool_call>` appear in the same deltafixed in v0.22.0
- day 77XGrammar #625featurefeat: expose normalize_tool_choice and unify Qwen XML structural tag buildersin v0.2.1
- day 76SGLang #24283bug[Bug] EAGLE/NEXTN speculative decoding crashes scheduler on response_format (xgrammar type mismatch)fixed in v0.5.11
- day 74XGrammar #610fixfix: align builtin structural tags with chat templates for reasoning and tool callsin v0.2.0
- day 72XGrammar #609fixrefactor: unify reasoning parameter and rename model keysin v0.2.0
- day 43vLLM #38603bug[Bug]: Streaming last chunk contains non-empty tool_calls with empty fields "type" causing type validation error open
- day 37vLLM #38106bug[Bug]: tool_choice="required" + speculative decoding with lukealonso/Qwen3.5-397B-A17B-NVFP4 leads to failed tool calls.open
- day 28vLLM #36774fix[Bugfix] fix Qwen3.5 tool calling bugin v0.18.0
- day 23vLLM #36769bug[Bug]: Qwen3.5 tool calling bugfixed in v0.18.0
- day 23vLLM #36794bug[Feature]: Improve Error Message When Max Length Reached With Tool Call=Requiredfixed in v0.18.0
- day 19vLLM v0.17.0supportmodel supportin v0.17.0
- day 8SGLang #19222bug[Bug] Anthropic API compatibility: Images in tool results are not correctly processed in SGLangfixed in v0.5.10rc0
- day 8SGLang v0.5.9supportmodel supportin v0.5.9
- day 8TensorRT-LLM v1.3.0rc5supportmodel supportin v1.3.0rc5
- day 0vLLM #34650bugBug: Speculative Decoding (MTP) Causes </think> Detection Failure in Structured Output + Reasoning Modefixed in v0.26.0
1open tool-calling issue upstreamvLLM 1 · SGLang 0 · TensorRT-LLM 0 · XGrammar 0 · oldest open 164 days
- 011 tool-calling bug is open upstream for Kimi K2.5 (vLLM 1), the oldest for 164 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix73 daysreport to release
Open now1oldest 164 days
Opened / fixed12 / 11fixed: in a release
First server supportday 2vLLM; SGLang day 28; TensorRT-LLM day 42
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (28)
- day 246vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 154vLLM #46610support[Frontend] Add Streaming Parser Engine and new Kimi k2.5/k2.6/k2.7 Parserin v0.25.0
- day 148SGLang #25071fixkimik2_detector fix the normal text detection before tool call. in v0.5.14
- day 148vLLM #46344fix[Frontend] Fix Kimi K2 tool call IDs for required tool choicein v0.24.0
- day 113vLLM #43203bug[Bug]: Not support role tool of imageopen
- day 100SGLang #23950fixfix(function_call): handle Kimi-K2.5 bare numeric tool call IDsin v0.5.12
- day 100SGLang #23953featurefeat(constrained): two-phase reasoning grammar + --enable-strict-thinkingin v0.5.12
- day 92vLLM #41182bug[Bug]: When using streaming tool calls in kimi-k2.5, only the content before the tool call can be obtainedfixed in v0.25.0
- day 91vLLM #41067bug[Bug]: KimiK2ReasoningParser silently corrupts streaming output when stop sequences buffer textfixed in v0.20.2
- day 84SGLang #23341bug[Bug] Kimi-K2 tool parser streaming content droppingfixed in v0.5.15.post1
- day 84SGLang #23363bug[Bug] KimiK2Detector streaming parser silently drops / hangs on multi-turn tool calls with long arguments (full rewrite proposed)fixed in v0.5.20
- day 82vLLM #38579fix[Bugfix] Kimi-K2 tool parser streaming - fix token leakage, argument truncation, and content droppingin v0.20.0
- day 74SGLang #21593fixFix tool call constrained decoding and parsing for models with native formatsin v0.5.11
- day 65SGLang #21948bug[Bug] `kimi_k2` tool-call parser produces invalid JSON escape sequences when streamingfixed in v0.5.10
- day 62SGLang #21666bug[Bug] kimi-k2.5 invalid message schemafixed in v0.5.13
- day 60vLLM #38441bug[Bug]: kimi-k2 tool parser regex is off a tiny bitfixed in v0.20.0
- day 51SGLang #19552fix[feat] Enhance Kimi-K2/K2.5 function call and reasoning detectionin v0.5.10rc0
- day 51vLLM #37438fix[Bugfix] Add Kimi-K2.5 reasoning/tool parser aliases and tool_call_id supportin v0.18.0
- day 50SGLang #20878bug[Bug] Improve Kimi-K2.5 tool parserfixed in v0.5.12.post1
- day 50XGrammar #562fixBetter kimi-k2 series models support in `builtin_structural_tag`in v0.1.33
- day 48SGLang #20688bug[Bug] [kimi k2.5] the content field was mixed with function callsfixed in v0.5.13
- day 48vLLM #37184bug[Bug]: Kimi-K2.5 Tool Parser Critical Issues - 87% Accuracy, 8KB Limit, Token Leakagefixed in v0.22.0
- day 42TensorRT-LLM v1.3.0rc7supportmodel supportin v1.3.0rc7
- day 28SGLang v0.5.9supportmodel supportin v0.5.9
- day 25SGLang #18996fixfix tool handling in OpenAIServingChatin v0.5.9
- day 6SGLang #18086bug[Bug] Kimi-k2.5 tool call parser may not work probabilistically.fixed in v0.5.10rc0
- day 2vLLM v0.15.0supportmodel supportin v0.15.0
- day 2SGLang #17914fixFix flaky tool calls in the Kimi K2.5 modelin v0.5.9
3open tool-calling issues upstreamvLLM 2 · SGLang 1 · TensorRT-LLM 0 · XGrammar 0 · oldest open 145 days
- 013 tool-calling bugs are open upstream for GLM 5 (vLLM 2, SGLang 1), the oldest for 145 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix72 daysreport to release
Open now3oldest 145 days
Opened / fixed9 / 6fixed: in a release
First server supportday 13SGLang; vLLM day 14; TensorRT-LLM day 41; XGrammar day 44
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (22)
- day 261TensorRT-LLM #21109fix[None][fix] GLM reasoning parser: drop the owed </think> after a tool call; read the thinking mode off the promptmerged, not released
- day 257SGLang #45143fix[Perf] function_call: stop rescanning unclosed GLM tool calls per chunkmerged, not released
- day 240SGLang #43273bug[Bug] Glm47MoeDetector streaming emits invalid JSON for array/object/boolean values that aren't strict JSON ("1, 2", "True", empty)open
- day 240TensorRT-LLM #19935fix[None][fix] GLM tool and reasoning parsers: streaming matches the whole-text parsein v1.4.0rc0
- day 224vLLM #58315bug[Bug]: GLM required tool-call grammar accepts split marker tokens that the streaming parser treats as contentopen
- day 214SGLang #39136fix[Fix] Preserve GLM tool argument types across JSON Schema unionsin v0.5.20
- day 132SGLang #28449supportExpand parser auto detection coveragein v0.5.14
- day 117vLLM #44843bug[Bug]: cc with vllm glm-5-fp8 function call erroropen
- day 109vLLM #44098bug[Bug]: GLM tool-call streaming final chunks repeat metadata and combine arguments with finish_reasonfixed in v0.25.0
- day 69TensorRT-LLM #13150support[TRTLLM-11999][feat] Add GLM-4.7/GLM-5 tool parserin v1.3.0rc13
- day 65vLLM #39870fix[BugFix] Support custom tool parsers when tool_choice is `required` and named functionin v0.19.1
- day 64SGLang #22595fixfix: normalize tool message content for GLM5.1 chat templatein v0.5.11
- day 62vLLM #39757bug[Bug]: GLM-5 tool calls in stream mode get error tool namefixed in v0.24.0
- day 53SGLang #22133bug[Bug] GLM-5 Eagle ValueError: Tokens not accepted: 154879fixed in v0.5.13
- day 44XGrammar v0.1.33supporttool-calling supportin v0.1.33
- day 43XGrammar #571support[feat] support glm47/glm5 tool call xml stylein v0.1.33
- day 41TensorRT-LLM v1.3.0rc9supportmodel supportin v1.3.0rc9
- day 29vLLM #36857bug[Bug]: The arguments invoked by the tool in the GLM-5 streaming output cannot be parsed into the JSON format.fixed in v0.22.0
- day 14SGLang #19345bugOpenAI-compatible tool-call stream can emit malformed JSON args (GLM-5 via NIM)fixed in v0.5.12
- day 14vLLM v0.16.0supportmodel supportin v0.16.0
- day 13SGLang v0.5.9supportmodel supportin v0.5.9
- day 1vLLM #34449bug[Bug]: GLM-5-FP8 malformed tool callsfixed in v0.19.1
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now0
Opened / fixed0 / 0fixed: in a release
First server supportday 13SGLang; vLLM day 23
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (4)
- day 153vLLM #57933support[Rust Frontend] Add MiMo V2.6 parser supportin v0.30.0
- day 23vLLM v0.21.0supportmodel supportin v0.21.0
- day 13SGLang v0.5.11supportmodel supportin v0.5.11
- day 8SGLang #23811support[Feature] Xiaomi MiMo-V2.5 day0 supportin v0.5.11
1open tool-calling issue upstreamvLLM 1 · SGLang 0 · TensorRT-LLM 0 · XGrammar 0 · oldest open 142 days
- 011 tool-calling bug is open upstream for GLM 5.1 (vLLM 1), the oldest for 142 days.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix15 daysreport to release
Open now1oldest 142 days
Opened / fixed10 / 9fixed: in a release
First server supportday 28SGLang
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (19)
- day 115SGLang #32522fix[Fix]: render tool_reference schema regardless of tool_result part orderin v0.5.17
- day 96SGLang #30925bug[Bug] GLM tool-call parser leaks raw <tool_call> markup into content when tool_choice="none"fixed in v0.5.15.post1
- day 77SGLang #28449supportExpand parser auto detection coveragein v0.5.14
- day 72vLLM #45915support[Frontend] Add Streaming Parser Engine and new GLM4.7/GLM5.1/GLM5.2 Parserin v0.24.0
- day 65vLLM #45273bug[Bug]: v0.19.1 /v1/responses fails for multi-turn Codex requests (GLM-5.1)open
- day 51vLLM #43851bug[Bug]: GLM chat template crashes with "'None' has no attribute 'items'" when tool_call arguments is the string "null"fixed in v0.22.1
- day 38SGLang #25388bug[Bug] JSON grammar abort with EAGLE/NEXTN + SPEC_V2 overlap in PD disaggregated decodefixed in v0.5.16
- day 35vLLM #42400bug[Bug]: GLM-5.1 tool call parsing fails intermittently when used as backend for Claude Codefixed in v0.24.0
- day 32vLLM #42167bug[Bug]: GLM5.1 tool call (with MTP) in streaming mode, arguments cannot be combined as a complete dictfixed in v0.21.0
- day 28SGLang v0.5.11supportmodel supportin v0.5.11
- day 15SGLang #22702featureSupport defer_loading field at function level for Chat Completions APIin v0.5.11
- day 10vLLM #39870fix[BugFix] Support custom tool parsers when tool_choice is `required` and named functionin v0.19.1
- day 9SGLang #22922bug[Bug] Tool-call-parser fails to format function calls for GLM5.1 (works for GLM5) in cursorfixed in v0.5.14
- day 9SGLang #22595fixfix: normalize tool message content for GLM5.1 chat templatein v0.5.11
- day 9vLLM #39899fix[bugfix] Normalize tool message content from array to string formatin v0.19.1
- day 6vLLM #39253fix[Bugfix] Fix GLM tool parser streaming with MTP or stream intervalin v0.19.1
- day 5vLLM #39611bugGLM-5.1-FP8: Tool results ignored via /v1/chat/completions with --tool-call-parser glm47 but work correctly via /v1/completionsfixed in v0.19.1
- day 5vLLM #39614bug[Bug] GLM-5.1-FP8: tool result content replaced with `<tools>\n</tools>` in /v1/chat/completions when --chat-template-content-format is autofixed in v0.20.0
- day 4vLLM #39574bug[Bug]: glm4_moe_tool_parser._is_string_type crashes on /v1/responses streaming with 'FunctionTool' object has no attribute 'function' (affects GLM-4.5/4.7/5.1)fixed in v0.22.0
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now0
Opened / fixed0 / 0fixed: in a release
First server supportday 13SGLang; vLLM day 23
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (4)
- day 153vLLM #57933support[Rust Frontend] Add MiMo V2.6 parser supportin v0.30.0
- day 23vLLM v0.21.0supportmodel supportin v0.21.0
- day 13SGLang v0.5.11supportmodel supportin v0.5.11
- day 8SGLang #23811support[Feature] Xiaomi MiMo-V2.5 day0 supportin v0.5.11
- 01Upstream has fixed all 8 tool-calling bugs reported for Llama 3.1 8B Instruct, 16 days from report to release at the median.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix16 daysreport to release
Open now0
Opened / fixed8 / 8fixed: in a release
First server supportday 0vLLM; TensorRT-LLM day 37
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (20)
- day 825vLLM #63569fix[Tool Calling] Migrate llama3_json parser to parser enginemerged, not released
- day 647XGrammar #614otherReduce FSM compilation overheadin v0.2.0
- day 647XGrammar #616fixperf: reduce structural-tag compile time at high tool countin v0.2.0
- day 490SGLang #13935support[misc] add llama3.1 chat templatein v0.5.6
- day 472SGLang #12431feature[router] Support structured model output for openai and grpc routerin gateway-v0.1.6
- day 443SGLang #11373fix[router][grpc] Fix tool call streaming bugs: empty tool names, state pollution, and panicsin v0.5.4
- day 442SGLang #11342fix[router][grpc] Add dependencies in Cargo.toml to support chat template renderingin v0.5.4
- day 431SGLang #10550featureUse jsonschema to constrain required or specific tool choicein v0.5.3
- day 385SGLang #9114fix[Bugfix] Fix accuracy-test-1-gpu failure caused by `builtin_tools`in v0.5.1
- day 240vLLM #15236bug[Bug]: Major issues with guided generation / structured output in vLLM (up to and including v0.8.1); many examples provided by vllm in /examples and structured_outputs.html doc do not workfixed in v0.8.3rc1
- day 200SGLang #3387bug[Bug] Streaming tool use is mixing in regular text with returned toolsfixed in v0.4.3
- day 195SGLang #3277bug[Bug] Error when performing function calling: `Exception: Extra data: line 1 column 41`fixed in v0.4.3
- day 175vLLM #12028bug[Bug]: another example of structured output xgrammar does not supportfixed in v0.7.0
- day 167vLLM #11774bug[Usage]: Running OpenAI Swarm with vLLM-hosted modelsfixed in v0.8.0rc1
- day 136XGrammar #122bug[Bug] Incorrect parsing for JSON Schemas with referencesfixed in v0.1.8
- day 91SGLang #1750bug[Bug] Llama 3.1/3.2 model in FC mode output continue past where it should stopfixed in v0.3.6
- day 66vLLM #8912bug[Bug]: Tool calling on Llama 3.1/3.2 fails with KeyError: '<tool_call>'fixed in v0.6.3
- day 66vLLM #8343support[Feature] Add support for Llama 3.1 and 3.2 tool usein v0.6.3
- day 37TensorRT-LLM v0.12.0supportmodel supportin v0.12.0
- day 0vLLM v0.5.3supportmodel supportin v0.5.3
8open tool-calling issues upstreamvLLM 1 · SGLang 4 · TensorRT-LLM 2 · XGrammar 1 · oldest open 124 days
- 018 tool-calling bugs are open upstream for DeepSeek V3.1 (vLLM 1, SGLang 4, TensorRT-LLM 2, XGrammar 1), the oldest for 124 days.
- 02SGLang v0.5.22 fixed #35565 ([Bug] 13 tool-call parsers silently delete the message when generation stops at a tool-call open marker) 12 days ago: providers on an older version still have the bug.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix19 daysreport to release
Open now8oldest 124 days
Opened / fixed15 / 7fixed: in a release
First server supportday 22SGLang; vLLM day 23; TensorRT-LLM day 140
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (42)
- day 431TensorRT-LLM #20989bug[Performance]: DeepSeek streaming tool parsers rescan the whole tool call on every increment (30 s of CPU for one 64k-character call)open
- day 431TensorRT-LLM #21004bug[Bug]: DeepSeek V3 and V3.1 streaming tool calls lose their arguments when the call ends in the last deltaopen
- day 424vLLM #60362fix[Bugfix][Tool Parser] Parse multi-line JSON arguments in the DeepSeek V3 and V3.1 parsersmerged, not released
- day 408SGLang #42373bug[Bug] DeepSeekV31Detector streaming drops the arguments when a complete tool call arrives in one deltaopen
- day 406SGLang #42131bug[Bug] DeepSeekV31Detector.structure_info() omits `<｜tool▁calls▁begin｜>`, so its own detect_and_parse returns no tool callopen
- day 406SGLang #42136bug[Bug] deepseekv31 detector (streaming) drops the text that shares a delta with the start of a tool callopen
- day 406SGLang #42137bug[Bug] deepseekv31 detector (streaming): greedy name regex drops a call and attaches the next call's arguments to itopen
- day 370TensorRT-LLM #17575fix[#17574][fix] Complete zero-argument tool calls in the streaming tool parserin v1.3.0rc25
- day 363SGLang #35565bug[Bug] 13 tool-call parsers silently delete the message when generation stops at a tool-call open markerfixed in v0.5.22
- day 363TensorRT-LLM #17903fix[#17580][fix] Stream the text that precedes a tool call in DeepSeek parsersin v1.3.0rc25
- day 362TensorRT-LLM #17573fix[#17572][fix] Emit the withheld buffer in DeepSeek streaming tool parsersin v1.3.0rc25
- day 356TensorRT-LLM #17572bug[Bug]: DeepSeek streaming tool parsers drop text they withheld while waiting on a tool-call tokenfixed in v1.3.0rc25
- day 356TensorRT-LLM #17580bug[Bug]: DeepSeek streaming tool parsers drop text that precedes a tool call in the same deltafixed in v1.3.0rc25
- day 321vLLM #48020bug[Bug]: DeepSeek-V3/V3.1 tool parser drops the tool call when start+end tokens land in the same streaming deltaopen
- day 314vLLM #46512other[Rust Frontend] Add error context in tool parser failuresin v0.25.0
- day 312XGrammar #680bug[Perf-Issue] cold-compile is slow for tool-calling tags with many toolsopen
- day 306SGLang #28449supportExpand parser auto detection coveragein v0.5.14
- day 233SGLang #21593fixFix tool call constrained decoding and parsing for models with native formatsin v0.5.11
- day 145vLLM #32240other[Refactor] [6/N] to simplify the vLLM openai chat_completion serving architecture (#32240)in v0.14.0
- day 140TensorRT-LLM v1.2.0rc7supporttool-calling supportin v1.2.0rc7
- day 132SGLang #13394fixFix DeepSeekV31's structural tag triggerin v0.5.7
- day 132TensorRT-LLM #10239other[None][chore] Unify DS tool parser namesin v1.2.0rc6.post1
- day 120TensorRT-LLM #10010support[TRTLLM-9604][feat] DS R1 & V3.1 tool parserin v1.1.0
- day 116vLLM #30675other[Refactor] [2/N] Move tool parsers into the vLLM main directory (#30675)in v0.13.0
- day 100vLLM #29693other[Misc] Refactor tokenizer interface (#29693)in v0.12.0
- day 85SGLang #11589fix[Tool Call] Steamline function arguments when tool_choice="auto" for deepseekv31_detectorin gateway-v0.1.6
- day 84SGLang #13163otherRemove EBNF Composer (#13163)in gateway-v0.1.6
- day 75vLLM #27974other[Refactor] Lazy import tool_parser (#27974)in v0.11.1
- day 70SGLang #12123fixFix DeepSeek chat templates to handle tool call arguments type checking (#11700)in v0.5.5
- day 62SGLang #11954bug[Bug] Function call not working in claude code after merging Anthropic API related issuefixed in gateway-v0.3.0
- day 56SGLang #11700bug[Bug] DeepSeekV3 series model chat_template contains redundant tojson in tool_calls, affecting Multi-Round FunctionCallfixed in v0.5.5
- day 37SGLang #10550featureUse jsonschema to constrain required or specific tool choicein v0.5.3
- day 34vLLM #25405fix[Bugfix] Fix DeepSeekV31ToolParser to correctly parse multiple tools in non-streaming outputin v0.11.0
- day 23vLLM v0.10.2supporttool-calling supportin v0.10.2
- day 22SGLang v0.5.2supporttool-calling supportin v0.5.2
- day 19SGLang #10209fixtool-call(dsv3): Fixed a parse problem when there are multiple function definitions in tool_callsin v0.5.2
- day 13SGLang #9895fixUpdate tool_chat_template_deepseekv31.jinjain v0.5.2
- day 12SGLang #9891bug[Bug] jinja2.exceptions.UndefinedError: 'dict object' has no attribute 'function'fixed in v0.5.2
- day 6SGLang #9446supportSupport DeepSeek-V3.1 tool callin v0.5.2
- day 4vLLM #23492fix[Fix] DeepSeek V3.1 tool parser error messagein v0.10.2
- day 3SGLang #9552bug[Bug] Function call not work with DeepSeek v3.1fixed in v0.5.2
- day 2vLLM #23454supportSupport DeepSeek-V3.1 tool callin v0.10.2
- 01Upstream has fixed all 1 tool-calling bugs reported for Mistral Nemo, 20 days from report to release at the median.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix20 daysreport to release
Open now0
Opened / fixed1 / 1fixed: in a release
First server supportday 4vLLM; SGLang day 6
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (5)
- day 240vLLM #14884bug[Bug]: if chat_template loaded from disk, jinja exception thrown from _try_extract_ast()fixed in v0.8.3rc1
- day 119vLLM #10333fix[Tool parsing] Improve / correct mistral tool parsingin v0.6.4
- day 60vLLM #8515support[Model] Add mistral function calling format to all models loaded with "mistral" formatin v0.6.2
- day 6SGLang v0.2.0supportmodel supportin v0.2.0
- day 4vLLM v0.5.3supportmodel supportin v0.5.3
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now0
Opened / fixed0 / 0fixed: in a release
First server supportday 5SGLang
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (1)
- day 5SGLang v0.5.19supportmodel supportin v0.5.19
13open tool-calling issues upstreamvLLM 7 · SGLang 5 · TensorRT-LLM 0 · XGrammar 1 · oldest open 103 days
- 0113 tool-calling bugs are open upstream for GLM 4.7 (vLLM 7, SGLang 5, XGrammar 1), the oldest for 103 days.
- 02SGLang v0.5.22 fixed #33324 ([Bug] Glm47MoeDetector sets tool_index to the tool's position in tools, not the call index (same pattern as #25073)) 12 days ago: providers on an older version still have the bug.
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix53 daysreport to release
Open now13oldest 103 days
Opened / fixed30 / 17fixed: in a release
First server supportday 2SGLang; vLLM day 29; TensorRT-LLM day 85; XGrammar day 95
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (78)
- day 312TensorRT-LLM #21109fix[None][fix] GLM reasoning parser: drop the owed </think> after a tool call; read the thinking mode off the promptmerged, not released
- day 312vLLM #64615fix[Bugfix][Frontend] Leave GLM-4.7 array and object tool arguments unpinnedmerged, not released
- day 311vLLM #64535fix[Bugfix][Parser] Parse closed GLM-4.7 tool arguments once while streamingmerged, not released
- day 308SGLang #45143fix[Perf] function_call: stop rescanning unclosed GLM tool calls per chunkmerged, not released
- day 307vLLM #63581bug[Bug]: Streamed tool_call `index` skips calls the parser rejects; the OpenAI SDK stream helper raises IndexError (GLM-4.7, MiniMax-M2, Kimi-K2)open
- day 305SGLang #45953bug[Bug] Glm47MoeDetector.detect_and_parse drops tool calls when the function name is followed by a newline (non-streaming)open
- day 305SGLang #46037bug[Bug][function_call] GLM tool-call parsers turn JSON null into the string "None"/"null"open
- day 302SGLang #41233fix[Fix] glm47: Disable tool grammar for `tool_choice="none"`merged, not released
- day 297vLLM #61328fix[Frontend] Let an unset `--tool-strict-level` fall back to the tool parser's defaultin v0.32.0
- day 294vLLM #57322feature[Frontend] Support CustomToolCall in the non-Harmony Responses APIin v0.32.0
- day 294vLLM #61253fix[Bugfix][Parser] Stop re-parsing the whole GLM tool argument per chunkin v0.32.0
- day 291SGLang #43273bug[Bug] Glm47MoeDetector streaming emits invalid JSON for array/object/boolean values that aren't strict JSON ("1, 2", "True", empty)open
- day 291TensorRT-LLM #19935fix[None][fix] GLM tool and reasoning parsers: streaming matches the whole-text parsein v1.4.0rc0
- day 290vLLM #60703bug[Bug]: glm47 streaming parser silently swallows the rest of the stream (including the real tool call) after a stray `<tool_call>` token in proseopen
- day 290vLLM #59627fix[Bugfix][Structured Output] Feed the token that implicitly ends reasoning to the grammarin v0.32.0
- day 289XGrammar #940feature[Python] Add opt-in token_markers to get_model_structural_tagmerged, not released
- day 288SGLang #42765bug[Bug] Glm47MoeDetector: a string parameter whose value is a quoted literal loses its quotesopen
- day 287vLLM #59879fix[Bugfix][Frontend] Build tool-call grammars from the tools the prompt rendersin v0.32.0
- day 284vLLM #56403fix[Frontend] Constrain non-strict GLM-4.7 tool calls with a shallow structural tagin v0.31.0
- day 284vLLM #59654fix[Bugfix][Rust Frontend] Preserve whitespace in GLM string argumentsin v0.31.0
- day 283vLLM #59608bug[Bug]: Structured output skips the `<tool_call>` that implicitly ends reasoning, so GLM-4.7/5.x `required`/named tool calls are lostfixed in v0.32.0
- day 282vLLM #59143other[Rust Frontend] Replay roundtrip output grammars through XGrammarin v0.31.0
- day 281XGrammar #939bugBuiltin structural tags admit control markers assembled from ordinary sub-tokensfixed day 289, not released
- day 281vLLM #59004fix[Rust Frontend] Relax schema-aware tool argument conversionin v0.31.0
- day 275vLLM #58315bug[Bug]: GLM required tool-call grammar accepts split marker tokens that the streaming parser treats as contentopen
- day 272vLLM #57826bug[Bug]: minimax_m2 / glm47_moe / deepseek_v4 arg converters drop the last parameter when the model omits the closing tag (same class as #57699)open
- day 268vLLM #56635fix[Parser] Fix: correct parser frontend handling of reasoning end and boundary tokensin v0.30.0
- day 267vLLM #56200fix[Refactor] Derive is_reasoning_end from the engine grammarin v0.30.0
- day 267SGLang #38890featureSupport non-strict GLM47 tool calls with EBNF constraintsin v0.5.20
- day 266XGrammar #892featurefeat: add a parallel_tool_calls param for builtin structural tagsin v0.2.7
- day 265SGLang #39136fix[Fix] Preserve GLM tool argument types across JSON Schema unionsin v0.5.20
- day 255vLLM #55080bug[Bug]: tool_choice='none' silently deletes tool-call-shaped contentopen
- day 249SGLang #36626fix[Fix] Resolve tool argument types through top-level anyOf/oneOf/allOfin v0.5.19
- day 248TensorRT-LLM #14290feature[TRTLLM-12758][feat] honor named function in tool_choice for non-harmony modelsin v1.3.0rc25
- day 247TensorRT-LLM #17575fix[#17574][fix] Complete zero-argument tool calls in the streaming tool parserin v1.3.0rc25
- day 240SGLang #35564bug[Bug] 8 tool-call parsers give wrong or missing calls in token-by-token streamingfixed day 310, not released
- day 227SGLang #33901bug[Bug] Glm47MoeDetector emits only the first of multiple tool calls in one final streaming incrementfixed in v0.5.22
- day 227SGLang #33902bug[Bug] GLM streaming parsers forward unknown tools when SGLANG_FORWARD_UNKNOWN_TOOLS is disabledfixed in v0.5.22
- day 224SGLang #33324bug[Bug] Glm47MoeDetector sets tool_index to the tool's position in tools, not the call index (same pattern as #25073)fixed in v0.5.22
- day 212vLLM #49412bug[Bug]: Streaming vs non-streaming content whitespace mismatch at tool-call boundaries (ParserEngine: qwen3/nemotron_v3/seed_oss/glm47_moe/gemma4)open
- day 210vLLM #49248bug[Bug]: glm47 tool parser silently returns empty arguments when the model omits the opening <arg_value> tagopen
- day 210XGrammar #704fixprevent XML whitespace state explosionin v0.2.5
- day 200XGrammar #695fixfix(structural_tag): fix several models with `tool_choice="required"` to terminate instead of looping.in v0.2.4
- day 199SGLang #30644bug[Bug] GLM tool-call parser corrupts string arguments that look like numbers with underscores (e.g. "123_456" -> "123456")fixed in v0.5.20
- day 183SGLang #28449supportExpand parser auto detection coveragein v0.5.14
- day 178vLLM #45915support[Frontend] Add Streaming Parser Engine and new GLM4.7/GLM5.1/GLM5.2 Parserin v0.24.0
- day 174SGLang #28149supportSupport GLM-4.7 function calling via structural tagsin v0.5.14
- day 162vLLM #44326bug[Bug]: GLM-4.7 streaming tool parser drops inline zero-argument tool callsfixed in v0.24.0
- day 156vLLM #43582fix[Rust Frontend] Add reasoning/tool parser & renderer roundtrip testsin v0.22.0
- day 138vLLM #42026fix[Bugfix] Preserve leading/trailing whitespace in GLM non-streaming tool parserin v0.20.2
- day 120TensorRT-LLM #13150support[TRTLLM-11999][feat] Add GLM-4.7/GLM-5 tool parserin v1.3.0rc13
- day 116vLLM #39870fix[BugFix] Support custom tool parsers when tool_choice is `required` and named functionin v0.19.1
- day 112vLLM #39253fix[Bugfix] Fix GLM tool parser streaming with MTP or stream intervalin v0.19.1
- day 110vLLM #39574bug[Bug]: glm4_moe_tool_parser._is_string_type crashes on /v1/responses streaming with 'FunctionTool' object has no attribute 'function' (affects GLM-4.5/4.7/5.1)fixed in v0.22.0
- day 108SGLang #20543fixfix: do not strip whitespace from GLM tool call values (#20543)in v0.5.11
- day 95XGrammar v0.1.33supporttool-calling supportin v0.1.33
- day 94XGrammar #571support[feat] support glm47/glm5 tool call xml stylein v0.1.33
- day 86vLLM #37386fixfix(glm47): improve tool call parsing and content normalizationin v0.18.0
- day 85vLLM #37277bug[Bug]: GLM47 Tool Call Bugfixed in v0.18.0
- day 85TensorRT-LLM v1.3.0rc8supportmodel supportin v1.3.0rc8
- day 81SGLang #20542bug[Bug] `Glm4MoeDetector` and `Glm47MoeDetector` strip significant whitespace from tool call argument valuesfixed in v0.5.11
- day 42vLLM #33218fix[Bugfix] GLM-4 tool parser: incremental string streamingin v0.15.1
- day 31SGLang #17608bug[Bug] GLM 4.7 Structured Decoding not working with "Invalid grammar request with cache hit" errorfixed in v0.5.8
- day 29vLLM v0.14.0supporttool-calling supportin v0.14.0
- day 25vLLM #32436bug[Bug]: GLM 4.7 Tool Call issuefixed in v0.19.1
- day 21vLLM #32213bug[Bug]: GLM4 tool parser crashes with TypeError when parsing tools with no argumentsfixed in v0.14.0
- day 18SGLang #15753fixFix GLM-4.7 MoE Detector complex JSON Schema type parsingin v0.5.8
- day 13SGLang #16371bug[Bug] GLM-4.7 Streaming Response Returns Invalid JSON Format for Multiple Tool Callsfixed in v0.5.10rc0
- day 8SGLang #15754fixFix: Handle empty func_name and None values in GLM MoE detectorsin v0.5.7
- day 7vLLM #31438fix[Bugfix] Preserve tool call id/type/name in streaming finish chunkin v0.14.0
- day 6vLLM #31443bug[Bug]: Tool name is lost in chat_completion_stream_generatorfixed in v0.14.1
- day 5vLLM #31437bug[Bug]: Streaming tool calls missing id/type/name in finish chunkfixed in v0.14.0
- day 4vLLM #31379bug[Bug]: GLM47 the empty parameter tool report TypeError: expected string or bufferfixed in v0.16.0
- day 2SGLang #15721bug[Bug] GLM 4.7 tool calling error in Claude Codefixed in v0.5.10rc0
- day 2SGLang gateway-v0.3.0supporttool-calling supportin gateway-v0.3.0
- day −1SGLang #15520support[model-gateway]: Tool parser for glm47in gateway-v0.3.0
- day −2SGLang #15333support[GLM-4.7] GLM-4.7 Tool Parser and Doc Updatein gateway-v0.3.0
- day −2vLLM #30876supportGLM-4.7 Tool Parser and Doc Updatein v0.14.0
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now0
Opened / fixed0 / 0fixed: in a release
First server support—no support change found
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (1)
- day 11vLLM #61727supportAdd Step-5 Preview support (#61727)merged, not released
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now0
Opened / fixed0 / 0fixed: in a release
First server supportday −109SGLang
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (2)
- day −51SGLang #34237fix[Fix] lfm2 detector: recover tool calls dropped by common model-outpu…in v0.5.19
- day −109SGLang v0.5.14supportmodel supportin v0.5.14
Upstream
Open tool-calling issues for the model in vLLM, SGLang, TensorRT-LLM and XGrammar, in days since it was listed on OpenRouter: +1 when a bug is reported, −1 when its fix is released (the repository's first release after the fix). Below, what happened on each server. Hover a marker for the issue, click to open it. Items are matched to the model by GPT-6 Luna reading them, so a few may be missing or misplaced.
Median time to fix—report to release
Open now0
Opened / fixed0 / 0fixed: in a release
First server supportday 22SGLang
open issuesbug opened, fixed sincebug still openfix releasedfix merged, not releasedfirst release supporting the model
All items (2)
- day 22SGLang v0.5.21supportmodel supportin v0.5.21
- day 6SGLang #38526supportAdd Ling-3.0-flash-VL model supportin v0.5.20