Mistral Large 4 Preview: benchmark results
Mistral Large 4 public preview, a natively multimodal 1T-parameter MoE with 52B active parameters, available through the API from October 6, 2026. Its own weights were not released: the open weights Mistral published on October 29 are an improved checkpoint after further reinforcement learning, listed separately as Mistral Large 4. Provider: Mistral. Released 2026-10-06. Access: API.
Aggregate Rating 1746 ± 8, rank #191 of 10870 rated models, from 373 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Mistral Large 4 Preview - Cybench | 93 | Challenges Solved (%) | 100 |
| Mistral Large 4 Preview - Dense200 (Bounding Box Detection) | 42 | Score (%) | 100 |
| Polish Knowledge Benchmark - Economy & Business | 84.39 | Accuracy (%) | 97.1 |
| Polish Knowledge Benchmark - Education & Science | 86.61 | Accuracy (%) | 97.1 |
| Vals AI SAFE-Teen - Relationship Boundaries | 19 | Failure rate (%), no teen-aware system prompt | 97.1 |
| AA GDP.pdf - Legal - Criterion Pass Rate | 86.43 | Criterion Pass Rate (%) | 96.8 |
| CyberGym-E2E-AA | 81.68 | Pass@1 (%) | 94.1 |
| Harvey LAB-AA - Corporate M&A | 97.11 | Criterion Pass Rate (%) | 93.8 |
| Polish Knowledge Benchmark - Ecology & Environment | 86.39 | Accuracy (%) | 92.6 |
| Polish Knowledge Benchmark - Law & Administration | 92.15 | Accuracy (%) | 92.6 |
| Polish Knowledge Benchmark - Media & Journalism | 81.58 | Accuracy (%) | 92.6 |
| Chatbot Arena (Text - Math) | 1486 | Arena Score | 92.4 |
Interactive version: theaggregate.ai/model?slug=mistral-large-4-preview · How It Works · Data refreshed daily, snapshot 2026-11-12.