MiniMax M2 — Benchmarks

Benchmark scores for MiniMax M2 aggregated from public leaderboards (7 boards), with how it ranks among open models and how it compares with its nearest neighbours. See hardware requirements for what you need to run it.

llmrun Composite Score →

104.0

90% range 99.1–109.0 · from 5 boards

CodingNot enough data yet
Agents104.690% range 100–110
MathNot enough data yet
ScienceNot enough data yet
ReasoningNot enough data yet

MiniMax M2 reasoning benchmarks

Multi-step logic and problem-solving — can the model think through a hard problem rather than recall an answer.

Vending-Bench 2
Vending-Bench 2, from Andon Labs, is a long-horizon agent benchmark in which the model runs a simulated vending-machine business for a year, handling suppliers, pricing and cash flow. The score is the final money balance in US dollars.
160.6
160.6
median 2376.8open leader 8313.8
Overall best: Gemini 4 Carbon (unspecified) 16484.2
19th percentile · #14 of 17 open · #64 of 71 all
#14 of 17
#64 of 71
Lab-reported
Epoch AI · Oct 2025

MiniMax M2 coding benchmarks

Writing, editing, and fixing real code — measured by whether the resulting program actually runs and passes tests.

SWE-bench Verified
SWE-bench Verified tests whether a model can resolve real GitHub issues from popular open-source Python projects, scored on the official swebench.com leaderboard as the percentage of human-validated issues actually fixed. It is the headline measure of practical, agentic software-engineering ability — where open-weight models like Qwen3-Coder, GLM-4.6, Kimi K2 and DeepSWE are now competitive with the frontier.
61.0%
61.0%
median 62.2%open leader 75.8%
Overall best: live-SWE-agent + Claude 4.5 Opus medium (20251101) 79.2%
47th percentile · #11 of 20 open · #73 of 162 all
#11 of 20
#73 of 162
Independent
SWE-bench · Nov 2025
Terminal-Bench
Terminal-Bench measures whether a model can complete real, end-to-end tasks in a command-line environment — running commands, editing files, and chaining steps — making it an agentic test of practical software skill.
30.0%
30.0%
median 31.7%open leader 52.4%
Overall best: GPT 5.5 (unspecified) 84.7%
47th percentile · #9 of 16 open · #43 of 57 all
#9 of 16
#43 of 57
Lab-reported
Epoch AI · Oct 2025
SWE-bench Bash Only
SWE-bench Bash Only runs the SWE-bench Verified issues through a minimal, single-tool bash agent — no specialised scaffolding — so the score reflects the model's own agentic coding ability rather than the harness around it. A cleaner, harder read on raw software-engineering skill.
61.0%
61.0%
median 55.4%open leader 63.4%
Overall best: Claude 4.5 Opus (high reasoning) 76.8%
88th percentile · #2 of 9 open · #25 of 48 all
#2 of 9
#25 of 48
Independent
SWE-bench · Nov 2025

Other results (1)leaderboard-only, not part of the llmrun Score

Leaderboard-only, not part of the llmrun Score

LMArena WebDev
Code Arena (formerly WebDev Arena) is Arena's head-to-head coding leaderboard: two models each build a real web app or UI from the same prompt, and people vote blind on which result is better. The rating shown is a relative Bradley-Terry score, not a percentage — compare it only to other models on this board.
1297.1
1297.1
median 1415.2open leader 1665.0
Overall best: Claude Opus 5.5 (Max) 1809.8
15th percentile · #47 of 55 open · #127 of 146 all
#47 of 55
#127 of 146
Arena
Arena · Oct 2026

MiniMax M2 knowledge benchmarks

Breadth and depth of factual knowledge across many subjects, paired with the reasoning to apply it.

MMLU-Pro
MMLU-Pro is a harder, cleaned-up successor to MMLU with ten answer choices and more reasoning-heavy questions across 14 subjects, measuring broad knowledge and reasoning together.
82.0%
82.0%
median 52.0%open leader 88.0%
Overall best: Gemini-3.1-Pro 91.2%
88th percentile · #18 of 146 open · #54 of 257 all
#18 of 146
#54 of 257
Independent
TIGER-Lab

MiniMax M2 instruction following benchmarks

How precisely a model obeys explicit, verifiable instructions — format, length, and content constraints — and, more broadly, how much people prefer its responses head-to-head in blind human voting.

LMArena Text
Arena (arena.ai, formerly LMArena / Chatbot Arena) ranks models by millions of blind, head-to-head human votes on real prompts. The Arena Score shown here is a Bradley-Terry rating computed with Arena's default style control, so higher means people preferred that model's answers more often — it's a relative rating, not a percentage, so compare it only to other models on this board.
1343.5
1343.5
median 1321.4open leader 1488.2
Overall best: Gemini 4 Carbon (High) 1527.5
57th percentile · #81 of 188 open · #221 of 408 all
#81 of 188
#221 of 408
Arena
Arena · Oct 2026

Score vs VRAM in the ≤192 GB tier

MiniMax M2
Other models
Best Score for its VRAM
x: VRAM (GB, log scale) · y: llmrun Score

#26 of 82 that fit ≤192 GB

Each dot is a model at the best quant that fits ≤192 GB. The line steps up to the best Score available at each amount of VRAM; a dot below it is beaten by a model that needs no more memory.

Full ≤192 GB ranking →

MiniMax M2 vs its neighbours

Same boards for the nearest open models per llmrun Score. The highest score in each row is marked.

Board	MiniMax M2this model	Qwen3.8 27Bbest in classQ4_K_M · 18 GB	MiniMax M2.1nearest by ScoreQ4_K_M · 140 GB	GLM 4.6nearest by ScoreQ4_K_M · 217 GB	Mistral Medium 3.5 128Bnearest by ScoreQ4_K_M · 83 GB
llmrun Score	104.0	122.0	106.2	103.1	102.1
SWE-bench Verified	61.0%	—	—	68.2%	—
Terminal-Bench	30.0%	—	36.6%	24.5%	—
SWE-bench Bash Only	61.0%	—	—	55.4%	—
MMLU-Pro	82.0%	—	88.0%	—	—
LMArena Text	1343.5	1435.6	1384.5	1424.3	1426.4
Compare head to head		Compare	Compare	Compare	Compare

Scores aggregated from public benchmark sources (Epoch AI: 2 boards, Arena: 2 boards, SWE-bench: 2 boards, TIGER-Lab: 1 board) — each linked from the benchmark pages. llmrun does not run these benchmarks. Percentiles and ranks are among open models in the llmrun catalogue; the median and leader marks refer to the same field.