Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
MiniMax M2.7: Open Source Self-Evolving AI Agent Achieving 56.22% on SWE-Pro Benchmark
MiniMax has just open sourced MiniMax M2.7, a groundbreaking self-evolving agent model that achieves 56.22% accuracy on the demanding SWE-Pro software engineering benchmark and 57.0% on Terminal Bench 2. This MoE architecture model not only matches top proprietary models like GPT-5.3-Codex but also actively participates in its own development, iterating autonomously to improve performance by 30%. Developers and AI researchers now have access to this frontier-grade, high-performance agent on Hugging Face, enabling faster, cost-effective deployment of complex AI agents for software engineering, professional office work, and multi-agent collaborations.
MiniMax M2.7’s real-world capabilities extend beyond code generation to production debugging, reducing incident recovery time to under three minutes, and handling substantial portions of reinforcement learning workflows autonomously. This open-source release could reshape how AI models evolve and collaborate on complex tasks, providing a powerful tool for developing sophisticated, autonomous AI systems. Explore the technical details and download the model weights to start building with this innovative agent today.