Mistral 7B Instruct v0.2 punches above its weight class — for a 7 billion parameter model, it handles instruction-following with surprising coherence and handles multi-turn conversations reasonably well. It tends to be direct and concise rather than verbose, though it can struggle with complex multi-step reasoning or nuanced tasks that larger models handle more gracefully. Its open Apache 2.0 license makes it a practical workhorse that runs on modest hardware.
| Benchmark | Score | Type | Recorded |
|---|---|---|---|
| GPQA Diamond | 3.5 | accuracy | 26d ago |
| MMLU-Pro | 19.1 | accuracy | 26d ago |
| MATH | 3.0 | accuracy | 26d ago |
| IFEval | 55.0 | accuracy | 26d ago |
| MuSR | 7.6 | accuracy | 26d ago |
| BBH | 22.9 | accuracy | 26d ago |