01Overview02Models03Rankings04Compare05Quiz06Guide

Model

MiniMax M3 NO.41

Open-weight multimodal model with native video input and 1M context.

CompareBack to the index

MiniMax · MiniMax · Open · Open weights

Specification

Fig.01 — MiniMax — MiniMax

CHINA

Context window
1.0Mtokens
Max output
66Ktokens
Input price
$0.30/ 1M
Output price
$1.20/ 1M
Cached input
$0.06/ 1M
Throughput
—
Class
Open
Modalities
Text, Image, Video
Weights
Open
Parameters
428B
License
MiniMax Community License
Knowledge cutoff
Jan 2026
Released
Jun 2026

Overview

MiniMax M3 is MiniMax's current flagship, an open-weight (MiniMax Community License) ~428B MoE (≈23B active) multimodal system with sparse/lightning attention giving cheap 1M-token context and rare native video input. It offers excellent value and strong long-context and multimodal breadth, with respectable coding, but does not match the Claude/GPT frontier for sustained autonomous agentic coding. A standout choice when video understanding, long context, and open licensing matter.

Strengths

  • Native video input (rare among LLMs)
  • Cheap 1M-token context via sparse attention
  • Open weights (MiniMax Community License)
  • Strong long-context and multimodal breadth

Trade-offs

  • Agentic coding reliability below the frontier
  • English writing slightly behind top labs
  • Smaller tooling ecosystem in the West
  • Vision/video quality trails dedicated multimodal flagships

Fit

Best for

  • Video and multimodal understanding
  • Cheap long-context processing
  • Self-hosted open deployments
  • High-volume multilingual tasks

Not ideal for

  • Long autonomous coding agents
  • Mission-critical frontier reasoning

Capabilities

Ten axes, normalized 0–100 and scored the same way across the whole catalog — so a 78 here means what a 78 means anywhere else on the bench.

Fig.02 — Capability radar

0 — 100

Reasoning
78
Coding
70
Math
72
Writing
74
Knowledge
80
Speed
76
Agentic
68
Vision
72
Multilingual
78
Long Context
88

Benchmarks

Public results, with independent third-party runs marked. Bars normalize percentages against 100 and Elo ratings against a 1500 ceiling.

Independent index

Artificial Analysis Intelligence Index

Composite of ~9–10 independent evals · Artificial Analysis

44/ 100

BenchmarkResult
Artificial Analysis Intelligence IndexIndependentGeneral · Artificial Analysis
44%
SWE-bench VerifiedIndependentCoding · SWE-bench
80.5%
SWE-bench ProIndependentCoding · SWE-bench Pro
59%
GPQA DiamondIndependentReasoning · Independent aggregators
93%
LiveBenchIndependentGeneral · LiveBench
70%
LMArena EloIndependentGeneral · LMArena
1455Elo

Alternatives

Models in roughly the same class — the ones worth weighing against this record.

Compare side by side
01DeepSeek-V3.2DeepSeekUltra-cheap open MoE with standout math and efficiency.02Qwen3.7-MaxAlibaba QwenAlibaba's proprietary agentic reasoning flagship with a 1M context.03Mistral Large 3Mistral AIEurope's flagship open-weight MoE — a strong, cheap, multilingual generalist.

Index

PreviousKimi K2 ThinkingNextClaude Fable 5
All models/EDUCATION