of gpt-oss-20b's structured outputs fail on Parasail (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 | Qwen3.6 35B A3Bfp8 | 0.09% | 1 of 8 | — | 1.4M |
| 2 | Mistral Nemofp8 | 0.15% | 1 of 3 | — | 65k |
| 3 | Kimi K3fp4 | 0.20% | 1 of 20 | — | 25k |
| 4 | GLM 5.3 Flashfp4 | 0.28% | 1 of 29 | — | 591k |
| 5 | Llama 3.3 70B Instructfp8 | 0.54% | 2 of 5 | SambaNova Turbo 0.16% | 101k |
| 6 | Gemma 4 31Bfp8 | 0.55% | 4 of 13 | ModelRun [by Modular] (fp4) 0.21% | 462k |
| 7 | Gemma 4 26B A4Bbf16 | 0.55% | 6 of 11 | Prime Intellect (bf16) 0.23% | 540k |
| 8 | GLM 5.2fp4 | 0.61% | 3 of 22 | Venice (fp8) 0.23% | 16k |
| 9 | DeepSeek V4.1 Flash 1017fp8 | 0.70% | 4 of 33 | Reka AI 0.40% | 121k |
| 10 | MiniMax M3.1 Flashfp8 | 0.85% | 3 of 8 | Nebius Token Factory (fp4) 0.14% | 13k |
| 11 | Qwen3 235B A22B Instruct 2507fp8 | 0.98% | 2 of 3 | Venice (fp8) 0.86% | 271k |
| 12 | Gemma 3 27Bfp8 | 1.01% | 2 of 4 | NextBit (fp8) 0.74% | 484k |
| 13 | Qwen3.5 397B A17Bfp8 | 1.37% | 5 of 7 | AtlasCloud (fp8) 0.65% | 21k |
| 14 | gpt-oss-120bfp4 | 1.41% | 9 of 18 | MARA 0.76% | 23k |
| 15 | Kimi K2.6int4 | 1.53% | 1 of 13 | — | 11k |
| 16 | Qwen3.8 27Bfp8 | 1.77% | 9 of 16 | NEAR AI (fp8) 0.25% | 32k |
| 17 | DeepSeek V4 Flash 0423fp8 | 1.78% | 6 of 13 | Alibaba Cloud Int. (fp8) 0.66% | 1.1M |
| 18 | DeepSeek V4.1 Flashfp8 | 2.56% | 22 of 28 | DeepInfra (fp8/flex) 0.35% | 150k |
| 19 | DeepSeek V4 Flash 0731fp8 | 2.74% | 14 of 18 | AtlasCloud (fp4) 0.16% | 1.1M |
| 20 | Qwen3.5-9Bbf16 | 3.26% | 3 of 6 | Prime Intellect (bf16) 0.30% | 163k |
| 21 | DeepSeek V4 Pro 0813fp8 | 3.61% | 12 of 13 | Cloudflare 0.39% | 1.5k |
| 22 | GLM 5.3fp8 | 4.00% | 16 of 28 | Together 0.25% | 7.1k |
| 23 | gpt-oss-20bfp4 | 6.43% | 5 of 8 | CoreWeave (fp4) 1.17% | 103k |
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 Qwen3.6 35B A3B, its best model.
models, where the settings tell OpenRouter to skip Parasail.
dotjson plugs into your inference stack, so every model you serve gets perfect structured outputs from day 0.
Fix Parasail's structured outputs↵