MiniMax has developed M2.7, a model that participated in its own development through a self-evolution process. The model ran unsupervised for over 100 rounds, analyzing failures and modifying its own code, resulting in a 30% performance improvement without human direction. This represents a fundamental shift in how AI models get built, with the model serving as both creator and product in its own training pipeline.

M2.7 demonstrates impressive capabilities across benchmarks, achieving a 66.6% medal rate on MLE Bench Lite and matching GPT-5.3-Codex with a 56.22% score on SWE-Pro. The model shows sustained improvement over long-running tasks, making it genuinely useful for production environments. It can handle complex software engineering scenarios, correlate monitoring metrics, and make SRE-level decisions about production incidents.

The model is available on HuggingFace with weights for local deployment, while NVIDIA offers free API access. Commercial use requires prior authorization from MiniMax and requires displaying "Built with MiniMax M2.7" in interfaces. Despite these restrictions, M2.7 represents one of the more capable agentic models currently available with public weights, particularly for developers and teams working with document processing and complex engineering tasks.