- HumanEval 0-shot 92.1
- LiveCodeBench (v6) 0-shot 30.3
- MultiPL-E average 0-shot 81.4
- ArenaHard 0-shot 97.1
- IfEval 0-shot 89.4
- Math500 Instruct 0-shot 91
- GPQA Diamond 5-shot CoT 57.1
- MMLU Pro 5-shot CoT 77.2
- Best of RULER 32K / RULER 128K 96
- MMMU 0-shot 66.1
- DocVQA 0-shot 95.3
- AI2D 0-shot 93.7
- ChartQA 0-shot 82.6
Mistral Medium 3
Mistral Medium 3 ranks first on Arena-Hard-Auto.
31
Benchmarks with a score
1
Ranked first, out of five or more models
3
In the top three
Mistral AI May 2025 Mistral Medium 3
Closed weights Mid-size
132.8
#151 of 256
Epoch Capabilities Index
Self-Reported Index 129.0 #283 of 404
- Humanity's Last Exam 4.52% #43 /47
- GPQA Diamond 59.5% #138 /229
- HiPhO 40.4% #24 /36
Where it ranks
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
Knowledge and language
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Arena-Hard-Auto | 97.1% | #1 of 39 | Lab model card | |
| RULER | 96% | #3 of 42 | Lab model card | |
| IFEval | 89.4% | #40 of 135 | Lab model card | |
| LMArena (Elo) | 1,387 | #149 of 335 | Official leaderboard | |
| MMLU-Pro | 77.2% | #98 of 167 | Lab model card | |
| LLM Confabulation Leaderboard | 21.9% | #43 of 58 | Official leaderboard | |
| Judgemark v2.1 | 58.5 | #32 of 42 | Official leaderboard | |
| Humanity's Last Exam | 4.52% | #43 of 47 | Official leaderboard | |
| Global MMLU Lite | 55.1% | #43 of 43 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MATH Level 5 | 81.6% | #28 of 94 | Epoch AI run | |
| Misguided Attention | 50.6% | #18 of 45 | Official leaderboard | |
| GPQA Diamond | 59.5% | #138 of 229 | Epoch AI run | |
| HiPhO | 40.4% | #24 of 36 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 32.2% | #143 of 212 | Epoch AI run | |
| MATH-500 | 91% | #50 of 70 | Lab model card | |
| DTBench | 62.3% | #134 of 180 | Official leaderboard | |
| SciArena | 885 | #31 of 34 | Official leaderboard | |
| FrontierMath | 0.35% | #64 of 70 | Epoch AI run |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| HumanEval | 92.1% | #7 of 101 | Lab model card | |
| MultiPL-E | 81.4% | #3 of 18 | Lab model card | |
| LiveCodeBench v6 | 30.3% | #89 of 90 | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BFCL V4 | 37.7% | #30 of 65 | Official leaderboard | |
| LisanBench | 130.7 | #108 of 133 | Official leaderboard | |
| Elimination Game | 2 | #52 of 56 | Official leaderboard |
Vision, audio and robotics
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Human Pathogen Capabilities Test v2 (HPCT v2) | 24.7% | #69 of 71 | Official leaderboard |
Its lab's numbers put it at 129.0, 3.8 points below its ECI
The Self-Reported Index fits one score to the 13 capability results Mistral AI reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 125.2 to 132.8. How we compute it.
- Arena-Hard-Auto 97.1
- RULER 96
- DocVQA 95.3
- AI2D 93.7
- HumanEval 92.1
- MATH-500 91
- IFEval 89.4
- ChartQA 82.6
- MultiPL-E 81.4
- MMLU-Pro 77.2
- MMMU 66.1
- GPQA Diamond 57.1
- LiveCodeBench v6 30.3
Its lab's number next to a leaderboard
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
| Benchmark | Lab says | Board says | Gap |
|---|---|---|---|
| GPQA Diamond | 57.1% GPQA Diamond 5-shot CoT | 59.5% Epoch AI run | −2.4 pts |
Mistral AI's models over time
Mistral AI's models in color, against every other model on the Epoch Capabilities Index, by release date.
Source: Epoch Capabilities Index, Epoch AI (CC BY 4.0)
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.