MiniMax open-sources M2 agent model
MiniMax has released open weights for M2, a model built for coding and agent work that the company says costs 8% of Claude Sonnet 4.5's price and runs about twice as fast.
MiniMax announced that it has open-sourced and launched MiniMax M2, a model it describes as "born for Agents and code." The complete weights are on Hugging Face, and SGLang and vLLM already support self-hosted deployment.
API pricing is $0.30 per million input tokens and $1.20 per million output tokens (¥2.1 and ¥8.4 RMB). The online inference service runs at about 100 tokens per second, which the company says is "rapidly improving." MiniMax puts that at 8% of Claude 4.5 Sonnet's price at nearly double the speed. A free trial has been extended to November 7, 00:00 UTC.
MiniMax says M2 is built for end-to-end development in tools such as Claude Code, Cursor, Cline, Kilo Code and Droid. It also says the model handles long chains of tool calls across a shell, a browser, a Python interpreter and MCP tools. The company credits the cost and speed to an "efficient design of activated parameters," but gave no parameter counts.
All performance claims come from MiniMax. The company says M2 is close to the best non-Chinese models on tool use and deep search, a little behind them on programming, and among the best Chinese models for coding. It also says M2 placed in the top five on Artificial Analysis's combined index of 10 benchmarks. MiniMax says it compared the model head-to-head with Claude 4.5 Sonnet and several open-source models on practical tasks, and that it will publish those test sets on GitHub this week. It says it will share the algorithmic details behind M2 "in due course."
MiniMax has also launched an M2-powered version of its MiniMax Agent product in China and upgraded the international version. The product has two modes: a fast Lightning mode and a Pro mode for complex tasks.