MiniMax M2.7

• 🧠 Self-evolving LLM that participates in optimizing its own agent scaffolding.
• 🆕 First M2-series model "deeply participating in its own evolution".
• 📏 Context window around 198K tokens for long-horizon work.
• 🔧 Built for coding, agentic tool use, and office productivity workflows.
• 🎯 Vendor reports 56.2% SWE-Pro and 57.0% Terminal Bench 2.
• 🌐 Capabilities include reasoning, function calling, and web search.
• 🏢 Released March 2026 by MiniMax; non-commercial license.

MiniMax

MiniMax is an AI company building generative models across multiple modalities, with a focus that spans both language understanding and audio creation. Their rapid release cadence in early 2026—delivering several new models within just a few months—reflects…

MiniMax M2.7 is a text-only large language model from MiniMax, released in March 2026, positioned for autonomous, real-world productivity through advanced agentic capabilities and multi-agent collaboration. It targets complex software engineering, long-horizon agentic tool use, live production troubleshooting, and office document generation across Word, Excel, and PowerPoint. The model supports reasoning, code-optimized workflows, function calling, and web search, with a context window of roughly 198K tokens.

Its defining feature is "self-evolution": MiniMax describes M2.7 as its first model that deeply participates in its own development, autonomously running 100-plus rounds of scaffold optimization to achieve a reported 30% improvement on internal programming benchmarks. On the vendor's own benchmarks, M2.7 scores 56.22% on SWE-Pro and 57.0% on Terminal Bench 2.

Compared with its same-family predecessor MiniMax M2.5, MiniMax reports significant gains in professional finance tasks—for instance, autonomously reading annual reports and earnings calls to build revenue models and produce research output.

M2.7 sits between M2.5 and the later MiniMax M3 in the lineage. The model is distributed under a modified, non-commercial MiniMax license.