of Llama 3.1 8B Instruct's tool calls fail on CoreWeave (bf16).
Updated October 29, 2026
Tool-call index
Not all models are served equally by all providers. On a typical model, the worst provider breaks tool calls 19× as often as the best. Find the best one for your agents.
| # | Model | Broken tool calls | Rank | Best on the model | Tool calls a day |
|---|---|---|---|---|---|
| 1 | Qwen3.8 27Bfp8 | 0.29% | 4 of 21 | Darkbloom (fp4) 0.07% | 94k |
| 2 | DeepSeek V4.1 Flashfp8 | 0.55% | 16 of 35 | Prime Intellect 0.06% | 1.1M |
| 3 | GLM 5.3fp8 | 0.66% | 18 of 42 | Amazon Bedrock 0.08% | 16k |
| 4 | GLM 5.2fp4 | 0.73% | 9 of 27 | Friendli 0.15% | 439k |
| 5 | DeepSeek V4 Flash 0731fp8 | 1.01% | 5 of 22 | Sail Research (fp4) 0.04% | 423k |
| 6 | DeepSeek V4 Pro 0813fp8 | 1.30% | 15 of 21 | Ionstream 0.04% | 68k |
| 7 | Gemma 4 31Bfp4 | 1.70% | 6 of 12 | ModelRun [by Modular] (fp4) 0.09% | 209k |
| 8 | Kimi K2.6fp4 | 1.83% | 12 of 14 | SiliconFlow (fp8) 0.06% | 69k |
| 9 | gpt-oss-120bfp4 | 2.16% | 3 of 15 | Groq 0.04% | 90k |
| 10 | GLM 5.3 Flashnvfp4 | 2.18% | 29 of 34 | NextBit (fp4) 0.00% | 1.5M |
| 11 | DeepSeek V3.1fp8 | 2.27% | 2 of 3 | SiliconFlow (fp8) 1.04% | 2.0k |
| 12 | Kimi K2.7 Codeint4 | 3.35% | 11 of 11 | NovitaAI (int4) 0.03% | 7.4k |
| 13 | Granite 4.2 8Bbf16 | 3.71% | 2 of 2 | DeepInfra (bf16) 2.00% | 7.8k |
| 14 | MiniMax M3fp4 | 4.57% | 9 of 11 | ModelRun [by Modular] (fp8) 0.27% | 61k |
| 15 | Llama 3.3 70B Instructfp16 | 4.65% | 2 of 5 | Groq 2.21% | 54k |
| 16 | Beam 1.0 501B A23Bfp4 | 6.52% | 3 of 3 | Together 1.01% | 21k |
| 17 | Nemotron 3.5 Lightningbf16 | 7.94% | 3 of 4 | Darkbloom (int4) 0.01% | 103k |
| 18 | Nemotron 3.5 Super VLnvfp4 | 8.36% | 8 of 8 | Together 0.66% | 27k |
| 19 | gpt-oss-20bfp4 | 13.45% | 5 of 7 | Darkbloom (fp8) 0.13% | 37k |
| 20 | Llama 3.1 8B Instructbf16 | 67.14% | 2 of 2 | Groq 1.00% | 8.6k |
Get updates when a new model drops
And which provider to use.
For inference providers
Tool calling doesn't have to break
The same provider can handle tool calls well on one model and struggle on the next.
on Qwen3.8 27B, its best model.
models, where the settings tell OpenRouter to skip CoreWeave.