MiniMax M3
MiniMax M3
Open weights tuned for software work, and the fastest decoding of anything in its class.
Updated
The read
M3 is a coding-first open model with an attention design that decodes dramatically faster than a conventional transformer of the same size, which is what makes it usable inside an agent loop rather than just on a benchmark. It sits near the top of verified software engineering scores while costing a fraction of the closed models around it. The licence is a community licence with commercial conditions, so read it before shipping.
Where it shines
- Decodes far faster than comparable open models
- Near the top of verified software engineering benchmarks
- Million-token context, which coding agents genuinely use
Worth knowing first
- Community licence carries commercial restrictions
- Weaker outside code and agent work than its headline scores suggest
Index score
84.5
Provisional. This score is read from the product's public capability, not from the one-prompt rebuild the fully reviewed entries went through, so treat it as a placing rather than a verdict.
- Rank
- 10 of 53
- Builds
- LLMs
- Category
- Coding
- Output
- Published weights, plus a cheap API
- Pricing
- Free tier
- Best for
- Coding agents that need speed as much as accuracy
Scores are our own editorial judgement, weighted craft 55, speed 10, control 15 and value 20.
Also worth a look
Full index- 04Claude Sonnet 5Claude Sonnet 5ProvisionalThe coding workhorse: near the top on real software tasks, at a working price.Coding modelPaid only87.9
- 18Step 3.7 FlashStep 3.7 FlashProvisionalAn Apache-licensed vision model built for coding agents, and it routes work rather than brute-forcing it.Coding modelFree tier81.6
- 43Mercury 2Mercury 2ProvisionalA diffusion model for text, and it is several times faster than anything built the usual way.Coding modelPaid only73.2
