MiniMaxmodel systemcard

MiniMax M2.7

#26of 27 · general

superseded by M3, kept for the version ladder — measured by Vals, Epoch and Arena; AA and LiveBench list only M3

205kcontext

MITlicense

2026-03-18released

Standing

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

General

6.0#26/27

Math

9.0#26/27

Coding

5.7#27/27

Health

3.1#25/26

Legal

58.7#6/26

Instruct

5.6#24/27

Agents

9.4#26/27

Professional

6.5#25/27

Hallucination

1.9#26/27

Value

5.2#27/27

Measured

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

GPQA Diamond86.6%#23

MMLU-Pro80.4%#22

AIME 202591.0%#5

SWE-bench V73.8%#24

LiveCodeBench79.9%#21

SciCode47.0%#21

CritPt0.6%#24

Terminal-Bench52.8%#22

Arena1417#23

Vals Index41.6#22

Epoch Capabilities Index145.9#22

Operations

1417arena elo · 51,372 votes

Series

MiniMax's models on the board, by release

MiniMax M2.52026-02-12 · 8MiniMax M2.72026-03-18 · 6MiniMax M32026-06-01 · 25

Neighbors

nearest capability fingerprints · deltas describe the neighbor

Nemotron 3 Ultra→ more factual · follows instructions better HY3→ stronger professional · more factual DeepSeek V4 Flash→ stronger professional Mistral Large 3→ stronger math Mistral Medium→ stronger math