MiniMax Releases M2.7 Self-Evolving Agent Model
MiniMax released M2.7 on April 11, 2026. According to MiniMax, the model is built to handle long-running agent tasks and to refine its own behavior during a session. The weights are open for download.
What MiniMax Says About the Design
MiniMax describes M2.7 as a self-evolving or self-improving agent. The company outlines a design that combines short-term memory for the current task, a self-feedback loop that reviews its own work, and a self-optimization step that updates the model's approach based on what worked. MiniMax also highlights sustained performance over long sessions. Readers should check the official model card for exact specs on session length and memory handling.
Benchmark Results
MiniMax reports strong scores for M2.7 on SWE-Pro, a software engineering benchmark, and Terminal Bench 2, a test of shell and command-line work. Specific numbers should be read directly from the release notes or the model card, as independent testing has not yet confirmed them.
Open Source Access
The model weights are published under an open license, and developers can run M2.7 locally or deploy it on their own infrastructure. This matters for teams that cannot send sensitive data to closed cloud APIs.
Why It Matters
Most AI agents restart from scratch every session. If M2.7 delivers on its self-improvement claims, it would be a step toward agents that get better within a single long run. Teams building coding agents, research assistants, and automation workflows should evaluate it on their own tasks before relying on the headline benchmark numbers.
Discussion
Sign in to comment. Your account must be at least 1 day old.