MiniMax Open Sources M2.7 Self-Evolving Agent Model Achieving 56.22% on SWE-Pro Benchmark

MiniMax has open-sourced its M2.7 self-evolving agent model on Hugging Face, achieving a 56.22% score on the SWE-Pro benchmark and 57.0% on Terminal Bench 2. The company simultaneously released MMX-CLI, a Node.js-based command-line interface that provides native access to the MiniMax AI platform's full suite of generative capabilities including image, video, speech, music, vision, and search. MMX-CLI is designed for use by both human developers working in terminals and AI agents running in platforms like Cursor, Claude Code, and OpenCode. The release represents a significant move toward open-source autonomous coding agents with integrated multimodal capabilities.

MMX-CLI Command-Line Interface Features and Capabilities

• ▪MMX-CLI provides access to vision capabilities.
• ▪MMX-CLI is a Node.js-based command-line interface.
• ▪MMX-CLI provides access to image generation capabilities.
• ▪MMX-CLI exposes the MiniMax AI platform's full suite of generative capabilities.

Integration with AI Agents and Developer Tools

• ▪MMX-CLI is designed for use by AI agents running in Claude Code.
• ▪MMX-CLI is designed for use by human developers working in a terminal.

Perspective of MiniMax

• ▪MMX-CLI enables AI agents to autonomously access MiniMax's multimodal generation capabilities through command-line operations.
• ▪MiniMax M2.7's 56.22% score on SWE-Pro represents competitive performance in software engineering tasks.
• ▪MiniMax's release of model weights on Hugging Face demonstrates a commitment to open-source AI development.

Perspective of AI development community

• ▪Integration with platforms like Cursor and Claude Code allows AI agents to leverage MiniMax M2.7's multimodal capabilities in existing workflows.
• ▪The Node.js foundation of MMX-CLI makes MiniMax's capabilities accessible to the large JavaScript developer ecosystem.
• ▪Open-sourcing MiniMax M2.7 provides developers with an alternative to proprietary coding agent models.

2 sources