MiniMax has open-sourced its AI model "M2.7." Weights can be downloaded from HuggingFace, and NVIDIA is providing a free API to allow users to test the model without hardware load. [MiniMax]
The company provided M2.7 with a programming skeleton and let it run for over 100 rounds in an unsupervised manner. The model analyzed its own failures, modified the code, and executed evaluations. It also autonomously decided which changes to retain.
As a result, performance improved by 30% without human instruction. MiniMax positions this not merely as a benchmark result, but as a new way of thinking about AI model construction. [MiniMax]
Specifically, the model was given access to 22 machine learning competition problems on MLE Bench Lite, and three 24-hour trials were conducted. In each round, the model generated memory files and reflected its self-criticism in the subsequent round.
In its best run, it earned 9 gold, 5 silver, and 1 bronze medals. The average medal rate across the three runs was 66.6%. This ranks third, following Opus 4.6 at 75.7% and GPT-5.4 at 71.2%. [MiniMax]
For developers, the score on SWE-Pro is critical. M2.7 recorded 56.22%, which is equivalent to GPT-5.3-Codex. It also recorded 76.5 on SWE Multilingual and 52.7 on Multi SWE Bench.
Furthermore, it achieved a score of 55.6% on VIBE-Pro, which is close to Opus 4.6. The company stated that if provided with complete project requirements, the model returns usable deliverables.
On GDPval-AA, it achieved an ELO of 1495. This is the highest value among open-source models, surpassing GPT-5.3 and trailing only Opus 4.6 and Sonnet 4.6. [MiniMax]
In terms of practical application, it supports multiple high-fidelity edits of Word, Excel, and PPT files. In a financial analysis task for TSMC, the model performed everything from reading annual reports to building prediction models and creating the final report; internal financial officers evaluated it as being usable as a first draft.