MiniMax M2.7 Open-Sourced: Unveils Self-Evolution Paradigm, Ascend Provides 0-Day Adaptation

On April 12, 2026, MiniMax officially open-sourced its self-deep iterative model MiniMax M2.7. Ascend AI basic software and hardware achieved 0-day adaptation for the model, supporting full-process inference deployment on Atlas 800 A3 and Atlas 800I A2 series products via vllm Ascend.

M2.7 can build complex Agent Harness, complete high-complexity productivity tasks through Agent Teams and Tool Search Tool, and update its memory to drive reinforcement learning, opening a new paradigm of self-evolution. Ascend optimizes via FlashComm sequence splitting and operator fusion to significantly improve throughput performance in heavy-load scenarios.

The model matches GPT-5.3-Codex in SWE-Pro accuracy in software engineering, ranks first in GDPval-AA among open-source models in professional office, and builds OpenRoom interaction system in interactive entertainment to enhance character consistency and dialogue ability.