MiniMax-M2.7

MiniMax-M2.7, launched in March 2026, is MiniMax’s first model to be deeply involved in its own iteration. M2.7 has a built-in Agent Harness framework, which can independently build complex skill sets, update its own memory, and drive iterative optimization through reinforcement learning, forming a closed loop of "model-driven model evolution". An internal version improved programming scaffolding performance by 30% in more than 100 rounds of self-optimization. On the software engineering benchmark SWE-Pro, M2.7 scored 56.22% and VIBE-Pro scored 55.6%, which is close to the Opus 4.6 level. M2.7 has significantly enhanced the complex editing capabilities of the three-piece Office suite, with a GDPval-AA ELO score of 1495, the highest among open source software. In the case of 40 complex skills (a single skill exceeds 2,000 Tokens), M2.7 maintained a skill compliance rate of 97%.

Context

204800 tokens

Max output

131072 tokens

Input

llm

Output

llm