Mistral Large 3: benchmark results
Mistral AI's Large 3, a 675B sparse MoE (41B active) flagship (December 2025). Provider: Mistral. Released 2025-12-02. Access: API.
Aggregate Rating 1613 ± 5, rank #401 of 1656 rated models, from 817 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| MOEVE - Summarization | 78 | Mean overall summarization score x 100 (0-100) over four Ger | 100 |
| ROK-FORTRESS | 28.4 | Δ_ling (pp) (self-reported) | 100 |
| SWE-PRBench (Config C) | 13.5 | Composite review score (0-1, scaled to 0-100): per pull requ | 100 |
| SpeechMap Compliance | 98.2 | % Requests Completed | 99.1 |
| Polish Knowledge Benchmark - Biographies | 90.12 | Accuracy (%) | 98.5 |
| AGC-Bench - ocw | 1.63 | Dataset z-score | 97.6 |
| UGI Leaderboard | 56.86 | UGI Score | 97.6 |
| Polish Knowledge Benchmark - Religion & Philosophy | 94.43 | Accuracy (%) | 97.1 |
| Grip on LLMs - Dutch Simplification - INT Duidelijke Taal | 43.15 | SARI (0-100) | 96.7 |
| Grip on LLMs - Dutch tinyMMLU | 69 | Accuracy (%; GP-IRT estimate from 100 tinyBenchmarks items, | 96.7 |
| HELM Arabic - Alghafa - Multiple Choice Grounded Statement Soqal Task | 96.67 | Exact Match (%) | 96.2 |
| StereoTales | 109 | Emissions (self-reported) | 95.5 |
Interactive version: theaggregate.ai/model?slug=mistral-large-3 · How It Works · Data refreshed daily, snapshot 2026-10-20.