MiniMax has released M2.7, a new artificial intelligence (AI) model that takes a step toward self-improvement in AI. The model actively participates in its own development by updating its memory, building complex skills, and refining its learning processes. This creates a cycle of self-evolution where the model helps improve itself with less human involvement.
M2.7 shows strong abilities in real-world software engineering tasks such as full project delivery, bug troubleshooting, code security, and machine learning. On the SWE-Pro benchmark, it achieved a score of 56.22 percent, close to leading models. It also performs well in end-to-end project work and understanding complex engineering systems. In professional office tasks, the model handles complex editing in Word, Excel, and PowerPoint with high accuracy and supports multi-round revisions. It maintains strong consistency when using over 40 complex skills.
Self-evolution through autonomous agent workflows
The development process demonstrates the potential for self-improvement. An internal agent harness built with M2.7 interacts across different projects, manages data pipelines, monitors experiments, analyzes logs, debugs code, and suggests improvements. In reinforcement learning experiments, the model autonomously runs iterative loops: it analyzes failures, plans changes, modifies code, evaluates results, and decides whether to keep or revert updates. In one case, this led to a 30 percent performance gain on internal tests after more than 100 rounds.
In low-resource tests, M2.7 participated in 22 machine learning competitions on a single graphics processing unit. Using short-term memory, self-feedback, and self-optimization, it improved results over time and achieved an average medal rate of 66.6 percent across three 24-hour runs. This shows the model can handle multiple stages of machine learning workflows with growing independence.
The capabilities extend to production debugging, where M2.7 reduces incident recovery time to under three minutes by combining monitoring data, causal reasoning, and code fixes. It also supports multi-agent collaboration, maintains high emotional intelligence, and enables interactive entertainment experiences. These features accelerate the shift toward AI-native organizations where models handle larger portions of development cycles.
Overall, M2.7 highlights a path where artificial intelligence systems can recursively improve their own harnesses, skills, and performance. "Hard Takeoff has started," says AI analyst and popular YouTuber Matthew Berman.