Model Drop

MiniMax-M2

October 27, 2025 2 min read

MiniMax-M2, which MiniMax released on October 27, 2025, is an open-weights 230B mixture-of-experts model with 10B active parameters, built for coding and agent work.

Model: MiniMax-M2 (MiniMaxAI/MiniMax-M2)
Maker: MiniMax
Ship date: October 27, 2025
Pricing: $0.30 / $1.20 per M tokens (MiniMax API, free until November 7, 2025); free weights
Context: 196,608 tokens
Weights: Open (MIT)
Where you can use it:
- MiniMax Agent, free for a limited time
- MiniMax API and open weights
When to reach for it:
- Cheap, fast coding agents like Claude Code
- Self-hosting a capable agent model
- Skip it if your app drops past thinking
How it compares:
- Priced at 8% of Claude Sonnet 4.5
- Open peer of GLM-4.6 and Kimi K2 0905
What makes it different:
- Only 10B of 230B parameters active
- Interleaved thinking between tool calls
Early reactions: Artificial Analysis called its index score a new high for an open-weights model, and Simon Willison called it a very promising model.
Read more: MiniMax M2 & Agent: Ingenious in Simplicity, model card

The benchmarks

MiniMax's launch table runs M2 with its interleaved thinking against Claude Sonnet 4.5, GPT-5 (thinking) and open peers. Rival scores with an asterisk in MiniMax's table come from each maker's own reports, and the rest MiniMax ran itself, so settings don't always match. Terminal-Bench is the original version run in Claude Code, Humanity's Last Exam is the text-only subset with search and Python tools, and "DeepSeek-V3.2" in MiniMax's table was the V3.2-Exp release out at the time.

MiniMax's numbers

SWE-bench Verified (real GitHub fixes): 69.4%. −7.8 under Claude Sonnet 4.5 (77.2%), −5.5 under GPT-5 (74.9%), +1.4 over GLM-4.6 (68%), +1.6 over DeepSeek V3.2-Exp (67.8%).
Terminal-Bench (agentic terminal tasks): 46.3%. −3.7 under Claude Sonnet 4.5 (50%), +2.5 over GPT-5 (43.8%), +1.8 over Kimi K2 0905 (44.5%), +5.8 over GLM-4.6 (40.5%).
SWE-bench Multilingual (fixes across languages): 56.5%. −11.5 under Claude Sonnet 4.5 (68%), −1.4 under DeepSeek V3.2-Exp (57.9%), +0.6 over Kimi K2 0905 (55.9%), +2.7 over GLM-4.6 (53.8%).
BrowseComp (hard web research): 44%. −10.9 under GPT-5 (54.9%), −1.1 under GLM-4.6 (45.1%), +3.9 over DeepSeek V3.2-Exp (40.1%), +24.4 over Claude Sonnet 4.5 (19.6%).
Humanity's Last Exam (with tools) (expert questions): 31.8%. −3.4 under GPT-5 (35.2%), +1.4 over GLM-4.6 (30.4%), +4.6 over DeepSeek V3.2-Exp (27.2%), +7.3 over Claude Sonnet 4.5 (24.5%).
τ²-Bench (customer-service agent tasks): 77.2%. −7.5 under Claude Sonnet 4.5 (84.7%), −2.9 under GPT-5 (80.1%), +1.3 over GLM-4.6 (75.9%), +6.9 over Kimi K2 0905 (70.3%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 18.6. +0.1 over GLM-4.6 (18.5), +2.0 over DeepSeek V3.2-Exp (16.6), −2.1 under Claude Sonnet 4.5 (20.7).