Mistral Large 4: benchmark results
The open-weight release of Mistral Large 4 on October 29, 2026, which Mistral describes as an improved model after further reinforcement learning on the October 6 public preview. Natively multimodal MoE with about 1T total and 52B active parameters; Mistral publishes FP8 weights, with an NVFP4 quantization in a separate repository. Provider: Mistral. Released 2026-10-29. Access: Open.
Aggregate Rating 1768 ± 8, rank #144 of 10516 rated models, from 325 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AI Coding Daily (OpenCode) - React-TS Code Quality | 18.5 | React-TS Code Quality (max 20) points, LLM-judged rubric sco | 100 |
| Mistral Large 4 Release - B3 Agent Security - Attack Resistance | 97.1 | Attack Resistance (%) | 100 |
| Mistral Large 4 Release - Cognition FrontierCode 2.0 | 19.7 | Score (%) | 100 |
| Mistral Large 4 Release - Cyber Refusal (JailbreakBench, StrongREJECT, AgentHarm) | 96.6 | Average Refusal Rate (%) | 100 |
| AA GDP.pdf - Legal - Criterion Pass Rate | 88.25 | Criterion Pass Rate (%) | 99.6 |
| Nejumi 4 - MT-Bench (Japanese) - Writing | 99.5 | Judge rating (1-10, x10) | 97.2 |
| OTIS Mock AIME 2024-25 | 99.44 | Accuracy (%) | 92.9 |
| AA-Omniscience Hallucination Rate | 29.81 | Hallucination Rate (%) | 92.6 |
| Nejumi 4 - Toxicity - Violation Categories | 54.76 | Criteria met (%) | 91.6 |
| AA-Briefcase - DOCX - Rubric Pass Rate | 57.4 | Rubric Pass Rate (%) | 91.3 |
| SecureBio BioTIER Permit | 99.9 | Benign prompts answered (%) | 90.8 |
| Nejumi 4 - MT-Bench (Japanese) - Coding | 100 | Judge rating (1-10, x10) | 90.1 |
Interactive version: theaggregate.ai/model?slug=mistral-large-4 · How It Works · Data refreshed daily, snapshot 2026-11-21.