Mistralmodel systemcard

Mistral Medium

#24of 27 · general

2508/3.5 collapsed

128Bparams

256kcontext

Modified MIT Licensenon-commercial

2026-04-29released

weights ↗

Standing

dot = this model · shaded = the field's middle 60% · track = full field

General

8.5#24/27

Math

41.8#14/27

Coding

27.0#19/27

Health

58.7#10/26

Analyst

20.0#9/11

Legal

46.4#15/26

Instruct

15.5#23/27

Agents

10.3#24/27

Professional

11.9#24/27

Multimodal

2.8#19/19

Hallucination

9.3#25/27

Speed

46.4#12/22

Value

18.9#25/27

Open

16.0#11/11

Measured

≈ imputed · amber dot = lab-self-reported among the sources

GPQA Diamond59.5%#27

HLE9.1%#24

MMLU-Pro75.3%#26

AIME 202564.3%#8

SWE-bench V72.0%#25

LiveCodeBench44.8%#25

SciCode39.6%#25

CritPt0.0%#26

Terminal-Bench44.8%#25

GDPval21.6%#21

Arena1427#21

AA Index30.4#22

Vals Index37.8#23

Epoch Capabilities Index142.6#24

Operations

$1.5 / $7.5$ per 1M tokens, in / out

$3blended

122tokens / sec

1427arena elo · 10,917 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 Large 3→ looser instructions · cheaper MiniMax M2.5→ weaker math · looser instructions HY3→ weaker math · cheaper Kimi K2→ stronger coding MiniMax M3→ weaker math · cheaper