System Card← leaderboard
Mistralmodel systemcard

Mistral Large 3

#27of 27 · general
675Bparams · MoE
Apache 2.0license
2025-09-01released

Standing

dot = this model · shaded = the field's middle 60% · track = full field
General
5.4#27/27
Math
30.6#20/27
Coding
19.0#24/27
Health
61.8#9/26
Legal
55.6#7/26
Instruct
2.3#26/27
Agents
2.2#27/27
Professional
2.5#26/27
Multimodal
5.2#18/19
Hallucination
10.5#24/27
Speed
38.3#15/22
Value
22.2#24/27
Open
19.8#10/11

Measured

≈ imputed · amber dot = lab-self-reported among the sources
GPQA Diamond68.0%#26Artificial AnalysisLLM StatsVals.ai
HLE4.2%#25Artificial Analysis
MMLU-Pro79.8%#24Vals.ai
AIME 202542.9%#9Vals.ai
SWE-bench V41.4%#27Vals.ai
LiveCodeBench44.9%#24LLM StatsVals.ai
SciCode36.2%#26Artificial Analysis
CritPt0.0%#25Artificial Analysis
Terminal-Bench10.5%#27Artificial AnalysisVals.ai
GDPval7.0%#22Artificial Analysis
Arena1415#24Arena
AA Index15.9#23Artificial Analysis

Operations

$0.50 / $1.5$ per 1M tokens, in / out
$0.75blended
39tokens / sec
1415arena elo · 52,509 votes

Series

Mistral's models on the board, by release
Mistral Large 32025-09-01 · 5Mistral Medium2026-04-29 · 9

Neighbors

nearest capability fingerprints · deltas describe the neighbor
Mistral Medium→ follows instructions better · pricierMiniMax M2.5→ weaker math · stronger codingHY3→ stronger agent · cheaperMiniMax M2.7→ weaker mathKimi K2→ stronger coding · follows instructions better