MiniMax-M1 is a large open-weight hybrid-attention model with long context support, RL-based training, and multiple deployment options. It provides function calling, a chatbot, and API access, with published benchmarks across multiple tasks.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
MiniMax-M1 is described as the world's first open-weight, large-scale hybrid-attention reasoning model. It uses a hybrid Mixture-of-Experts (MoE) architecture with a lightning attention mechanism and is trained with large-scale reinforcement learning on diverse tasks. It claims a context length of 1 million tokens and reports reduced test-time FLOPs compared to some baselines (e.g., 25% of the FLOPs at a generation length of 100K tokens versus DeepSeek R1).
How it works
The model combines MoE with a hybrid-attention design to improve efficiency in RL scaling. It mentions two thinking budgets (40K and 80K) for model variants and introduces CISPO, an RL algorithm that clips importance sampling weights instead of token updates. It references evaluation across mathematics, coding, software engineering, tool use, and long-context tasks, with results reported against several benchmarks.
Getting started
Deployment options include vLLM for serving, or deploying with Transformers directly. Provided deployment guides are:
- vLLM Deployment Guide (./docs/vllm_deployment_guide.md)
- MiniMax-M1 Transformers Deployment Guide (./docs/transformers_deployment_guide.md)
Model download pages mentioned for the two thinking budgets:
- MiniMax-M1-40k
- MiniMax-M1-80k
Also, a Chatbot and an online API are available for general use and evaluation, along with a MiniMax MCP Server with media generation capabilities.
Recent releases
RELEASES (latest 0): - none
Traction
- Stars: 3164
- Forks: 283
- Open issues: 31
Behind the repo
Links to related projects include MiniMax-MCP and platform/integration tools; the repo links to Hugging Face pages and an arXiv preprint (Paper: 2506.13585).
Caveats
Licensing is Apache-2.0. The README discusses substantial model claims and benchmark results but does not provide versioned release notes beyond the noted latest releases being none. It also references multiple external deployment guides and APIs.