Mistral 7B Instruct (v0.2): benchmark results
Mistral AI's instruct fine-tune of Mistral-7B-v0.2, extending context to 32K tokens versus v0.1's 8K, with an increased rope theta and no sliding-window attention. Provider: Mistral. Released 2023-12-11. Access: Open.
Aggregate Rating 1370 ± 8, rank #1333 of 1656 rated models, from 389 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Hallucinations Leaderboard | 51.63 | Average Task Score (score) | 94.7 |
| Open LLM Leaderboard v1 - TruthfulQA MC2 | 68.26 | MC2 (%) (0-shot) | 87.4 |
| AlpacaEval 1.0 | 92.78 | Win Rate (%) | 84.2 |
| AlpacaEval | 82.98 | Length-Controlled Win Rate (%) | 82.5 |
| OpenCompass Language - Sentiment Analysis - English (CompassBench 2404) | 73.3 | Score (%) | 80 |
| SpeechMap Compliance | 81.5 | % Requests Completed | 79.2 |
| Open Chinese LLM - TruthfulQA MC | 58.78 | Accuracy (%) | 78.4 |
| Open PL LLM - PolEmo2-IN (multiple choice, 5-shot) | 83.24 | Accuracy (%) | 77.9 |
| OpenCompass Reasoning - Deductive - English (CompassBench 2404) | 18.6 | Score (%) | 76.7 |
| EconLogicQA | 0.32 | Accuracy | 76.5 |
| CyberMetric | 78.4 | 500-question accuracy (%) | 75 |
| RenoBench - Issue | 94.3 | Recall (%) of gold JATS issue annotations reproduced by the | 75 |
Interactive version: theaggregate.ai/model?slug=mistral-7b-instruct-v0-2 · How It Works · Data refreshed daily, snapshot 2026-10-20.