MiniMax M2.7: Open-Source AI Agent Matches Top Proprietary Models

Chinese AI company MiniMax has open-sourced M2.7, a new AI agent model that matches the best proprietary models on real-world software engineering tasks — and you can download it right now.

What Makes M2.7 Different

Most AI coding benchmarks test models on tidy algorithm puzzles. M2.7 was built for messier, more realistic work: debugging production code, analysing server logs, reviewing code for security flaws, and managing machine learning workflows.

On SWE-Pro — a benchmark that mirrors real software engineering scenarios — M2.7 scores 56.22%, matching OpenAI’s GPT-5.3-Codex. On Terminal Bench 2, which tests command-line problem solving, it hits 57.0%. Among open-source models, it holds the top ELO rating on GDPval-AA, trailing only Claude Opus 4.6, Sonnet 4.6, and GPT-5.4 overall.

The Self-Evolving Angle

Here is the part that caught researchers’ attention. M2.7 is the first model to actively participate in its own training. During development, MiniMax let the model run over 100 autonomous rounds of scaffold optimisation — essentially letting it fine-tune how it approaches tasks. The result was a 30% performance improvement driven by the model itself, not just by human engineers tweaking parameters.

MiniMax says this points toward a future where models can continuously improve with less human intervention.

A Licensing Wrinkle

There is one thing to watch. Shortly after the initial release, MiniMax quietly changed the license terms, drawing criticism from the open-source community. If you plan to use M2.7 commercially, check the current license on Hugging Face before building anything on top of it.

What This Means for You

If you use AI for coding — whether through Claude Code, Cursor, or other tools — M2.7 is worth paying attention to. An open-source model performing at the level of GPT-5.3 means more competition, lower prices, and more options for developers who want to run models locally without sending code to the cloud.

You can try M2.7 through Ollama for local use or via API providers like OpenRouter. For context on how it compares to other coding AI tools, see our Claude Code guide. And if you are just getting started with AI-assisted development, visit our beginner’s guide.