MiniMax has open-sourced MiniMax M2.7, its most capable Mixture-of-Experts model to date, which actively participated in its own development cycle. The model achieves state-of-the-art performance on real-world software engineering benchmarks, scoring 56.22% on SWE-Pro and 57.0% on Terminal Bench 2.

  • MiniMax M2.7 is the first model to autonomously optimize its own scaffold, running over 100 rounds of iterative improvement that yielded a 30% performance gain.
  • It handles 30–50% of MiniMax’s internal reinforcement learning workflows end-to-end and reduces production incident recovery time to under three minutes.
  • The model ranks as the highest open-source option on GDPval-AA with an ELO score of 1495, surpassing GPT-5.3 in professional office tasks.
  • It demonstrated strong autonomous ML experimentation capabilities on MLE Bench Lite, achieving a 66.6% average medal rate across three trials.

The release makes frontier-grade agentic capabilities freely accessible for developers to deploy and build upon.