Now a legacy model, it still suits developers wanting cheap API coding help or free open weights to run themselves.

Kenning's take

What it does

MiniMax M2.5 is a large language model from the Chinese company MiniMax, trained with reinforcement learning across hundreds of thousands of real-world environments. MiniMax reports 80.2% on SWE-Bench Verified and says it plans a project like a software architect before writing code. It is available through the MiniMax API, in MiniMax's own agent product, and as downloadable weights on Hugging Face; MiniMax now lists it as a legacy model after M2.7 and M3.

  • 80.2% on SWE-Bench Verified (MiniMax figure)
  • Trained on 10+ programming languages
  • Search and tool calling for agent tasks
  • Faster -highspeed API variant
  • Open weights on Hugging Face

Who it suits. Developers building coding agents and automated workflows who want a low-cost model they can also self-host.

Key facts

Category
AI models
Made by
MiniMax
Based in
China
Released
12 February 2026
How to get it
through the MiniMax API (pay as you go or a coding plan), in the MiniMax Agent app, and as open weights on Hugging Face under a modified MIT licence
Works on
API, Web, self-hosted
Free plan
Open weights are a free download
API price
$0.30 per million input tokens, $1.20 per million output; highspeed variant $0.60 input, $2.40 output

FAQ

When was MiniMax M2.5 released?

MiniMax announced M2.5 on 12 February 2026.

Is MiniMax M2.5 open source?

Its weights are on Hugging Face under a modified MIT licence.

How much does the MiniMax M2.5 API cost?

$0.30 per million input tokens and $1.20 per million output tokens.

Facts from the maker's own pages and the sources listed. Checked 2026-10-11.