MiniMax M2.7: Open-Source AI That Optimized Its Own Training

On April 12, MiniMax released M2.7, a 230-billion-parameter Mixture-of-Experts model with open weights. What sets it apart is that it was an active participant in its own training rather than a passive recipient of gradient updates: M2.7 was given write access to its persistent memory, the ability to build Python-based skills used during training, and permission to modify parts of the agentic training scaffold itself, including evaluation logic and loop detection. Using these capabilities, the model optimized its own sampling parameters, sharpened workflow instructions, and added logic to catch circular reasoning patterns that were wasting compute — yielding a documented 30% improvement in reinforcement-learning training throughput.

Architecturally, M2.7 is an MoE model that activates only about 10 billion of its 230 billion total parameters per token, giving it the inference cost profile of a much smaller model while retaining broad knowledge capacity. Deployment guides are available for SGLang, vLLM, and Hugging Face Transformers, with a minimum recommended setup of four GPUs at 96GB VRAM, scaling to eight GPUs for a 3-million-token context.

Benchmark results back up the claims: M2.7 scores 56.22% on SWE-bench Pro, matching GPT-5.3-Codex, and 57.0% on Terminal Bench 2. These scores place it among the top open-weight coding agents available, and because it's open-weight, enterprises can fine-tune and self-host it — something closed API models don't allow. For engineers, the significance lies in M2.7 being the first production model to turn 'self-evolving AI' from a research concept into a measurable training-efficiency gain.