Ran it, or it didn’t happen.
Benchmark open models on your own hardware. Runs count once their event log checks out.
Run your first benchmarkRecent verified runsWhereSpeed
gemma-3-12b-it Q4_K_M@li-wei-evals · Nov 6, 2026RTX 4090 · 24 GB89.3 tok/s
MiniCPM5-2B Q4_K_M@li-wei-evals · Nov 6, 2026L40S · 48 GBCloud297.2 tok/s
Qwen3-Coder-30B-A3B-Instruct Q4_K_M@saoirse-dev · Nov 6, 2026RTX 4090 · 24 GB224.4 tok/s- Qwen3-8B Q4_K_M@ben-hardware · Nov 6, 2026RTX 4090 · 24 GB141.5 tok/s
Phi-3.5-mini-instruct Q4_K_M@alexruns · Nov 6, 2026RTX 4060 Ti · 16 GB89.3 tok/s
PassedFailedNot supported
Cloud leaderboard
Every model runs on the same pinned machine. 17 models, 70 runs.
- Suite
- [email protected]
- Machine
- NVIDIA L40S · 48 GB
- Engine
- ollama 0.40.14
| # | Model | Pass rate |
|---|---|---|
| 1 | Qwen3.5-35B-A3BQ4_K_M | 96% |
| 2 | Qwen3.5-27BQ4_K_M | 93% |
| 3 | Qwen3-30B-A3B-Instruct-2507Q4_K_M | 92% |
| 4 |  GLM-4.7-FlashQ4_K_M | 89% |
| 5 | Qwen3-Coder-30B-A3B-InstructQ4_K_M | 89% |
| 6 | Qwen3-8BQ4_K_M | 89% |
| 7 | Qwen3-14BQ4_K_M | 87% |
| 8 | gemma-3-27b-itQ4_K_M | 86% |
| 9 |  gpt-oss-20bMXFP4 | 85% |
| 10 | gemma-3-4b-itQ4_K_M | 83% |
| 11 | Qwen3.5-2BQ4_K_M | 81% |
| 12 | Qwen3-4BQ4_K_M | 78% |
| 13 | gemma-3-12b-itQ4_K_M | 75% |
| 14 |   Llama-3.1-8B-InstructQ4_K_M | 71% |
| 15 | Phi-3.5-mini-instructQ4_K_M | 64% |
| 16 | MiniCPM5-2BQ4_K_M | 52% |
| 17 |   Mistral-7B-Instruct-v0.2Q4_K_M | 47% |
Ranked by checks passed. A task a model can’t do is shown, not counted as a zero.
Open model leaderboard
143 open-weight models in creative writing ranked by LMArena score · updated Oct 28, 2026
Show all 143 models
Scores: LMArena, CC BY 4.0. Measured on served versions, not local builds. A model with several served variants shows its best.