AI BriefingMiniMax Launches 1 Million-Token Context Model, Promises M3 Weight Release
Key point
MiniMax has launched the M3 model featuring a 1 million-token context and multimodal capabilities, with weights to be released within 10 days.
Details
MiniMax has launched M3, a model supporting a 1 million-token (1M) context window and native multimodal input. It is currently available first via API, with model weights and a technical report to be released via Hugging Face and GitHub within 10 days.
M3 aims for performance optimized for coding agent tasks, achieving the following results on major benchmarks:
- SWE-Bench Pro: 59.0%
- Terminal-Bench 2.1: 66.0%
- MCP Atlas: 74.2%
To efficiently handle long contexts, the model applies MiniMax Sparse Attention (MSA) technology. MSA selects relevant KV (Key-Value) blocks through pre-filtering, reducing per-token computation by a factor of 20 compared to the previous generation in a 1 million-token context environment.
The API provides endpoints compatible with OpenAI and Anthropic, and supports image and video input. Pricing is around $0.60 per million input tokens and $2.40 per million output tokens. Meanwhile, following the announcement of MiniMax's plans to list on Shanghai's STAR Market, its stock fell 16% on the Hong Kong stock exchange.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.