MiniMax
MiniMax

MiniMax M3

Release Date 2026-06-01(Knowledge cutoff 2026-01)1.0M IN · 512K OUTProprietary ModelAPI available

MiniMax M3 understands text, images and video and returns text for software development, long-document analysis and tool workflows. MiniMax describes broader visual and long-context capabilities than its earlier text-focused models, with sparse attention improving compute efficiency; savings in one operation do not imply equal reductions in total bills or task time. AIB rates document work and writing in the middle band at the time checked, supporting reviewable source-based reports and technical drafts. Official weights use the MiniMax Community License, so check its terms and hardware requirements rather than assuming standard MIT conditions. Available input length can depend on the serving route and configuration, while code execution and file edits require a tool environment and checks on the resulting changes.

API cost

Market price (median)
$0.29IN$1.15OUT
Price range5 providers
$0.23~$0.55IN$0.96~$2.2OUT
IN
OUT
11/2111/22
SourceOpenRouterLast updated

Real-world usage

No.474 places0.2% share · 472B tokens/wk
6.5T8/31
2.0T9/7
1.5T9/14
1.4T9/21
1.2T9/28
1.3T10/5
1.4T10/12
1.4T10/19
612B10/26
672B11/2
612B11/9
472B11/16
SourceOpenRouterLast updated

Output Speed

Averageper 2,000 tokens
67tok/s
First Output 0.7stotal: 30.5s
SourceOpenRouterLast updated

User Comparisons

Writing Style
Answer length
Average
bottom 41%
Vocabulary diversity
Plainer side
bottom 31%
Sentence rhythm
Steadier side
bottom 39%
Language mixing
High
top 25%
List usage
Often
20% of lines
Heading usage
Occasionally
12% of lines
Reasoning share
Heavier side
32% of total
Lies (Hallucination)
Honest answers
100%
Correct 0Honest don't-know 1Lies 0
Response Speed
instability ±125%
First response 9.7sTotal 22.1s
54 tok/s
12.9ms per token
User Reviews
  • •

    Accurately finds various sources and summarizes them properly. The Korean is natural too.

  • •

    Sometimes thinks way too much...

SourceAIB Inc.Last updated

Expert Performance Evaluation

18.9143467.3

More about AA Intelligence Index v5

15.3%$2,15888.9%74.2%22.3%

More about τ³-Banking

83.0%144337.7%48.3%68.1%20.6%29.8%46.3%143776.254.3%4.2%

More about Long Context Reasoning

143383.5%

More about Arena Text Factuality

2.0%149240.764059.0%65.2%42.4%58.6148447.1%80.5%14.7%148268.247.6%0.9%52.6%

More about Terminal-Bench 4.0

1410144782.9%1392142676.8114087.3%139857.5

More about Arena Writing, Literature & Language

39.0%3.7%142392.9%71.1%77.018.0%55.0%1458144274.545.8%17.0%8.0%78.9%33.7

More about Humanity's Last Exam

124878.1%

More about Arena Vision

Specialized fields

13611366145614671450143297.6%82.7%43.8%80.6%85.3%79.0%75.2%

More about Arena Korean

OddBench

64.0%Reasoning effort: Extra High65.1%25330

By effortNone 63.0%Extra High 64.0%

More about BullshitBench v2

See each evaluation’s detail page for its source and last updateData corrections & removal requests