MiniMax M2.7
Billed as MiniMax's first self-evolving model. Autonomous optimization over 100+ rounds is said to have yielded a 30% gain on an internal version.
MiniMax describes recursive self-improvement, in which the model updated its memory, built skills and eval sets, and modified its own scaffold. 229B parameters; SWE-Pro 56.22%, Terminal Bench 2 57.0%, MLE Bench Lite 66.6% medal rate. Weights followed in April.
- Date
- Wednesday, 18 March 2026
- Lab
- MiniMax
- Kind
- open-weights
- Access
- open weights (restricted license)
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-Pro | 56.22% | company |
| Terminal Bench 2 | 57.0% | company |
| MLE Bench Lite medal rate | 66.6% | company |
| GDPval-AA Elo | 1495 | company |
API launch 2026-03-18; HF repo created 2026-04-09 so weights landed after launch. 'Self-evolution' is a company description and not independently verified.
Sources
- www.minimax.io/news/minimax-m27-en
- huggingface.co/MiniMaxAI/MiniMax-M2.7
- platform.minimax.io/docs/release-notes/models
This record was checked against its sources on 6 October 2026. How we check