MiniMax M2.7 (Tencent Cloud)

MiniMax M2.7 is a reasoning-focused large language model from the Shanghai-based MiniMax team, positioned in their text lineup alongside the sibling M3 and M2.5 releases. Its defining characteristic, as described in independent reporting, is participation in its own training loop: M2.7 and its predecessors actively drive an internal reinforcement learning research harness, with the outlet estimating that 30–50% of the RL workflow that shapes the model is automated by the model family itself. This makes M2.7 a concrete example of recursive self-improvement tooling moving from theory into production, rather than a conventional static foundation model release.

M2.7 is marketed as an agent-oriented backend intended to power developer tools and AI assistants such as Claude Code, Kilo Code, and OpenClaw, with an explicit aim of reaching parity with leading U.S. models on reasoning benchmarks while keeping cost and latency attractive for enterprise workflows. The strategic shift is notable: MiniMax, known as one of China's more visible open-source-oriented AI players, is using M2.7 to show that closed-weights, agent-optimized variants can coexist with self-evolving research infrastructure. Practical fit centers on teams building multi-step coding or tool-use agents who want a reasoning model whose training pipeline itself is treated as a first-class product surface.