MiniMax M1 40K
M1 40K is one of two separately trained M1 reasoning releases. It builds on Text-01 and combines a mixture-of-experts network with lightning attention.
A different reinforcement-learning method
M1 uses CISPO, which clips importance-sampling weights rather than token updates. MiniMax trained the 40K and 80K versions with different thinking budgets. The converted Hugging Face format is part of the same release.MiniMax M1 40K model card
Choose a budget from the task
Practical guidance. Compare this release with M1 80K on tasks that have a clear correct result. Longer reasoning can consume more resources without fixing a poor plan. Record answer quality and completion delay together.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | MiniMax-M1-40kMiniMax-M1-40k-hf |
| Context window | 1,000,000 tokens Native M1 context capacity; 40K refers to the training-time thinking budget. MiniMax M1 40K model card |
| Size | 456 billion total parameters, with 45.9 billion active per token.MiniMax M1 40K model card |
| Thinking budget | Trained with a 40,000-token thinking budget. This is distinct from its input context capacity.MiniMax M1 40K model card |
| Licence | Apache 2.0.MiniMax M1 40K model card |
Related models
- MiniMax Text-01
MiniMax · MiniMax-01
A long-context language model that combines lightning attention, softmax attention and a mixture-of-experts network.
Launch date Downloadable - MiniMax M1 80K
MiniMax · MiniMax M1
A Text-01-derived reasoning model trained with an 80,000-token thinking budget and tool-call support.
Launch date not confirmedDownloadable