Wait Which Model?
← Back to directory
MiniMax

MiniMax-M2

Released Oct 27, 2025 · knowledge cutoff unpublished

Status
Superseded
Location
China
Modality
Text
Context window
205K
Max output
—
Speed
95 tok/s · 1.69s to first answer token
Price ($/MTok in / out)
$0.30 / $1.20
Cost per task
—
Open weights
Yes
Benchmarks5 of 8 reported

MMLU-Pro

82%

GPQA Diamond

78%

SWE-bench Verified

69.4%

AIME (latest)

78%

Humanity's Last Exam

12.5%

Terminal-Bench 2.1

—

LMArena Elo

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • Interleaved thinking alternates plan, act and reflect within a single response rather than reasoning once and executing blind — makes long, tool-heavy agent loops easier to debug
  • Genuinely fast for its price bracket — MiniMax reports roughly double the throughput of Claude Sonnet at a small fraction of the cost
  • Sparse MoE with only 10B active parameters keeps latency low on long agentic sessions despite the 230B total size

Weaknesses

  • Programming ability trails the leading closed models, competitive mainly against other open-weight coding models
  • Reliable tool use depends on the interleaved-thinking output format; wrapping it in a non-standard harness reportedly degrades results
  • As the first release in a fast-iterating series, MiniMax had already shipped two further point updates within four months

Officially announced 2025-10-27 (minimax.io/news/minimax-m2). Hugging Face model card lists 230B total / 10B active parameters. GPQA-Diamond 78, AIME25 78, HLE 12.5% (without tools, 31.8% with), MMLU-Pro 82 and SWE-bench Verified 69.4 are from the official Hugging Face README. A 46.3 'Terminal-Bench' figure also appears there without a version label, so left null here rather than assume 2.1. Context window 204,800 tokens per MiniMax's own API docs; max output isn't stated in MiniMax's docs (a third-party 131,072 figure was not independently corroborated, so left null). Licensed under MiniMax's 'Modified MIT' terms: commercial use is unrestricted, with an in-product attribution requirement triggered only above 100M monthly active users or $30M annual revenue.

For developers

API model strings

  • minimaxMiniMax-M2

Licence

MiniMax Modified MIT License · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor

News