of GLM 5.2's structured outputs fail on DeepInfra (fp4).
Updated October 31, 2026 · How we measure
Structured-output index
The same model breaks your pipeline's structured outputs 15× more often on some inference providers than on others. Find the best one for your pipelines.
Give this to your agent ↓
| # | Model | Broken structured outputs | Rank | Best for this model | Structured outputs a day |
|---|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flashfp8/flex | 0.35% | 1 of 28 | — | 254k |
| 2 | DeepSeek V4 Pro 0423fp8 | 0.36% | 1 of 8 | — | 79k |
| 3 | Gemma 4 26B A4Bfp8 | 0.40% | 3 of 11 | Prime Intellect (bf16) 0.23% | 375k |
| 4 | DeepSeek V4.1 Flashfp8/priority | 0.40% | 2 of 28 | DeepInfra (fp8/flex) 0.35% | 60k |
| 5 | GLM 5.3 Flashfp4 | 0.47% | 4 of 29 | Parasail (fp4) 0.28% | 618k |
| 6 | Gemma 4 31Bfp8 | 0.51% | 3 of 13 | ModelRun [by Modular] (fp4) 0.21% | 305k |
| 7 | DeepSeek V4.1 Flashfp8 | 0.52% | 4 of 28 | DeepInfra (fp8/flex) 0.35% | 2.2M |
| 8 | Kimi K3mxfp4 | 0.52% | 2 of 20 | Parasail (fp4) 0.20% | 12k |
| 9 | DeepSeek V4.1 Flash 1017fp8/flex | 0.65% | 2 of 33 | Reka AI 0.40% | 132k |
| 10 | DeepSeek V4 Flash 0731fp8 | 0.65% | 4 of 18 | AtlasCloud (fp4) 0.16% | 3.4M |
| 11 | DeepSeek V3.2fp4 | 0.67% | 1 of 5 | — | 254k |
| 12 | Qwen3.6 35B A3Bfp8 | 0.72% | 3 of 8 | Parasail (fp8) 0.09% | 87k |
| 13 | Qwen3.5 397B A17Bfp8 | 0.89% | 3 of 7 | AtlasCloud (fp8) 0.65% | 25k |
| 14 | DeepSeek V3.1fp4 | 0.90% | 3 of 4 | MARA 0.12% | 147k |
| 15 | DeepSeek V4.1 Flash 1017fp8 | 0.90% | 7 of 33 | Reka AI 0.40% | 678k |
| 16 | DeepSeek V4 Flash 0423fp8 | 0.96% | 2 of 13 | Alibaba Cloud Int. (fp8) 0.66% | 1.7M |
| 17 | Qwen3.5-9Bbf16 | 0.96% | 2 of 6 | Prime Intellect (bf16) 0.30% | 89k |
| 18 | Qwen3 235B A22B Instruct 2507fp8 | 1.00% | 3 of 3 | Venice (fp8) 0.86% | 346k |
| 19 | Gemma 3 27Bfp8 | 1.08% | 3 of 4 | NextBit (fp8) 0.74% | 254k |
| 20 | gpt-oss-120bbf16 | 1.08% | 4 of 18 | MARA 0.76% | 473k |
| 21 | DeepSeek V4 Pro 0813fp8 | 1.09% | 5 of 13 | Cloudflare 0.39% | 13k |
| 22 | DeepSeek V4.1 Flash 1017fp8/priority | 1.10% | 14 of 33 | Reka AI 0.40% | 417k |
| 23 | GLM 5.3fp4 | 1.21% | 9 of 28 | Together 0.25% | 7.4k |
| 24 | Qwen3.8 27Bbf16 | 1.41% | 7 of 16 | NEAR AI (fp8) 0.25% | 76k |
| 25 | MiMo-V2.6-Flashfp8 | 1.42% | 3 of 8 | GMICloud (bf16) 0.21% | 62k |
| 26 | Mistral Nemofp8 | 1.55% | 2 of 3 | Parasail (fp8) 0.15% | 246k |
| 27 | Nemotron 3.5 Lightningbf16 | 1.77% | 2 of 5 | CoreWeave (bf16) 1.75% | 18k |
| 28 | Gemma 4 31Bfp4/turbo | 2.21% | 12 of 13 | ModelRun [by Modular] (fp4) 0.21% | 1.7M |
| 29 | gpt-oss-20bbf16 | 2.26% | 2 of 8 | CoreWeave (fp4) 1.17% | 45k |
| 30 | Gemma 4 31Bfp8/priority | 2.45% | 13 of 13 | ModelRun [by Modular] (fp4) 0.21% | 94k |
| 31 | MiMo-V2.6-Profp8 | 2.59% | 1 of 3 | — | 88k |
| 32 | Nemotron 3.5 Super VLfp4 | 2.74% | 5 of 6 | Venice (fp4) 1.45% | 30k |
| 33 | Llama 3.3 70B Instructfp8/turbo | 3.48% | 5 of 5 | SambaNova Turbo 0.16% | 248k |
| 34 | Kimi K2.6fp4 | 3.64% | 10 of 13 | Parasail (int4) 1.53% | 25k |
| 35 | GLM 5.2fp4 | 4.64% | 17 of 22 | Venice (fp8) 0.23% | 79k |
Get updates when a new model drops
And which provider to use.
For inference providers
Structured outputs don't have to break
The same provider can handle structured outputs well on one model and struggle on the next. Implementations diverge on strings, optionals, composition and conditionals (the April announcement): the schema one provider enforces, another ignores.
on DeepSeek V4.1 Flash, its best model.
models, where the settings tell OpenRouter to skip DeepInfra.
dotjson plugs into your inference stack, so every model you serve gets perfect structured outputs from day 0.
Fix DeepInfra's structured outputs↵