MiniMax · model passport

MiniMax: MiniMax M3

minimax-m3

Open weights
MiniMax Community License licence
427.0B params
1,048,576 context
Released 2026-06-02(HF repo created)

Overall · LMArena: 1,441#84 of 395
Best percentile: 99.4IFBench (Artificial Analysis)
Cohort standing: 60.4median 57.3

What the data says

Benchmarked on 14 of 27 tracked axes. Every figure below is the model's own published result on that board.

Strongest axes

99.4 pct — IFBench (Artificial Analysis) (82.9%)
86.4 pct — GPQA Diamond (90.9%)

Weakest axes

10.9 pct — Agent Arena (-0.079)

Uneven: 88.5 percentile points between its best and worst axis — 99.4 pct on IFBench (Artificial Analysis) versus 10.9 pct on Agent Arena.

IFBench (Artificial Analysis)99.4 pct82.9%Verify →
GPQA Diamond86.4 pct90.9%Verify →
Coding80.8 pct1,497Verify →
Instruction following80.8 pct1,435Verify →
Hard prompts80.5 pct1,464Verify →
LiveBench79.1 pct67.3%Verify →
Overall79 pct1,441Verify →
Vision Arena65.4 pct1,236Verify →
WebDev Arena59 pct1,482Verify →
AIME (math)52.5 pct71.1%Verify →
Factuality Arena46.2 pct1,435Verify →
SimpleBench41.9 pct45.8%Verify →
Document Arena15.8 pct1,433Verify →
Agent Arena10.9 pct-0.079Verify →
Search ArenaNot benchmarked
FrontierMathNot benchmarked
SimpleQA VerifiedNot benchmarked
ARC-AGI-2Not benchmarked
Humanity's Last ExamNot benchmarked
BFCL v4 (function calling)Not benchmarked
SWE-bench Verified (live board)Not benchmarked
Terminal-Bench 4.0Not benchmarked
MMLU-Pro (Artificial Analysis)Not benchmarked
LiveCodeBench (Artificial Analysis)Not benchmarked
HHEM Hallucination (Vectara)Not benchmarked
SWE-bench VerifiedNot benchmarked
Terminal-BenchNot benchmarked

Percentiles are the model's standing within each board's own population. Axes marked "not benchmarked" have no published result for this model — that is an absence of data, not a low score.

How to run it

Run it yourself

RepositoryMiniMaxAI/Minimax-M3

Parameters427.0B

ArchitectureMiniMaxM3SparseForConditionalGeneration

LicenceMiniMax Community License

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "MiniMaxAI/Minimax-M3"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

Ollama

We don't sync Ollama tags, so we won't print a command we can't verify — a wrong one would pull the wrong model. Search the Ollama library →

Unsloth fine-tuning recipes are model-specific and we don't derive them here. Unsloth docs →

Use it hosted

API model IDminimax/minimax-m3

Context length1,048,576 tokens

Max output512,000 tokens

Acceptstext, image, video

Input price$0.30 per 1M tokens

Cached input price$0.06 per 1M tokens

Output price$1.20 per 1M tokens

API parameters this endpoint supports

Frequency penaltyInclude reasoningLogit biasLogprobsMax tokensMin-pPresence penaltyReasoningRepetition penaltyResponse formatSeedStop sequencesStructured outputsTemperatureTool choiceTool callingTop-kTop logprobsTop-p

Example request

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"minimax/minimax-m3","messages":[{"role":"user","content":"Hello"}]}'

View on OpenRouter →

Identifiers, limits and prices are synced from the upstream source shown above. Any field we haven't synced is left out entirely rather than guessed.

The minimax-m3 model from minimax has a shrunk composite score of 60.4 and a raw composite score of 62.7, with coverage on 14 of 27 axes.
standingVerify →
The minimax-m3 model, from minimax, has a standing of 60.4, compared with a cohort median standing of 57.3 across a cohort of 244 models over a 180-day window.
cohortVerify →

Every claim is a computed fact from the current board data, phrased by a low-temperature model call and rejected by a validation gate if any number is not in the source row. How the leaderboards work →