of gpt-oss-120b's structured outputs fail on CoreWeave (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 | Gemma 4 26B A4Bbf16 | 0.41% | 4 of 11 | Prime Intellect (bf16) 0.23% | 275k |
| 2 | DeepSeek V3.1fp8 | 0.43% | 2 of 4 | MARA 0.12% | 64k |
| 3 | MiniMax M3fp4 | 0.61% | 1 of 3 | — | 54k |
| 4 | Gemma 4 31Bfp4 | 0.68% | 5 of 13 | ModelRun [by Modular] (fp4) 0.21% | 811k |
| 5 | Llama 3.3 70B Instructfp16 | 0.73% | 4 of 5 | SambaNova Turbo 0.16% | 31k |
| 6 | GLM 5.3 Flashnvfp4 | 1.05% | 14 of 29 | Parasail (fp4) 0.28% | 1.1M |
| 7 | gpt-oss-20bfp4 | 1.17% | 1 of 8 | — | 542k |
| 8 | Qwen3.6 35B A3Bfp8 | 1.46% | 6 of 8 | Parasail (fp8) 0.09% | 48k |
| 9 | DeepSeek V4 Flash 0731fp8 | 1.63% | 8 of 18 | AtlasCloud (fp4) 0.16% | 1.1M |
| 10 | GLM 5.2fp4 | 1.64% | 9 of 22 | Venice (fp8) 0.23% | 150k |
| 11 | Nemotron 3.5 Super VLnvfp4 | 1.66% | 2 of 6 | Venice (fp4) 1.45% | 16k |
| 12 | Nemotron 3.5 Lightningbf16 | 1.75% | 1 of 5 | — | 117k |
| 13 | Beam 1.0 501B A23Bfp4 | 2.10% | 2 of 2 | Nebius Token Factory (fp4) 1.41% | 7.7k |
| 14 | Kimi K2.6fp4 | 2.24% | 3 of 13 | Parasail (int4) 1.53% | 52k |
| 15 | Qwen3.8 27Bfp8 | 2.50% | 10 of 16 | NEAR AI (fp8) 0.25% | 8.5k |
| 16 | DeepSeek V4.1 Flashfp8 | 3.02% | 25 of 28 | DeepInfra (fp8/flex) 0.35% | 952k |
| 17 | DeepSeek V4 Pro 0813fp8 | 4.69% | 13 of 13 | Cloudflare 0.39% | 13k |
| 18 | GLM 5.3fp8 | 5.29% | 21 of 28 | Together 0.25% | 3.9k |
| 19 | gpt-oss-120bfp4 | 6.53% | 16 of 18 | MARA 0.76% | 145k |
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 Gemma 4 26B A4B, its best model.
models, where the settings tell OpenRouter to skip CoreWeave.
dotjson plugs into your inference stack, so every model you serve gets perfect structured outputs from day 0.
Fix CoreWeave's structured outputs↵