Weights are available on Hugging Face; MiniMax describes M2.7 as its most capable open-source model and first to join its own development cycle.
AI Quick Take
- Weights published on Hugging Face let the community access and inspect M2.7.
- MiniMax reports 56.22% on SWE‑Pro and 57.0% on Terminal Bench 2 - benchmark figures, not deployment claims.
MiniMax has open-sourced MiniMax M2.7 and published the model weights on Hugging Face. The release positions M2.7 as MiniMax's most capable open-source model to date and identifies it as the company's first model to "actively participate in its own development cycle." The announcement highlights reported benchmark results of 56.22% on SWE‑Pro and 57.0% on Terminal Bench 2.
MiniMax describes M2.7 as a self‑evolving agent model that took part in its own development process; the company frames that participation as a change in how the model was iterated. Beyond the characterization of the model and the headline benchmark numbers, the source preview does not provide additional technical documentation in this announcement text.
Publishing weights on Hugging Face makes M2.7 available for inspection, replication and downstream work by researchers and practitioners; the reported scores provide a baseline for comparison but are presented as benchmark results in the release. The available information is concise, so readers wanting deeper technical detail or reproducibility evidence will need to consult the Hugging Face upload or follow-up materials from MiniMax.