MiniMax is a Chinese artificial-intelligence company, backed by Alibaba and Tencent, known for large language models built for agentic workflows and end-to-end coding, including the open-source MiniMax-M2 and the proprietary, self-evolving M2.7.
MiniMax released MiniMax-M2 as an open-source model on October 27, 2025, under a permissive MIT license. It is a sparse mixture-of-experts model with 230 billion total parameters and roughly 10 billion active per inference, a context window near 204,800 tokens, and native support for shell, browser, Python-interpreter, and 📝Model Context Protocol (MCP) tooling. It scores 69.4 on SWE-bench Verified, and benchmark aggregator Artificial Analysis has ranked it among the strongest open-source models for math, science, instruction-following, coding, and agentic tool use.
Cost is central to M2's positioning: MiniMax reports it runs at roughly 8% of the cost of comparable frontier systems — about 92% cheaper than Claude Sonnet 4.5 — while delivering faster inference, which has made it a popular provider for agentic and coding harnesses.
MiniMax later introduced M2.7, a proprietary model it calls "self-evolving": the company reports M2.7 was used to build its own research-agent harness and now autonomously handles 30% to 50% of its reinforcement-learning research workflow, including log analysis, debugging, and metric evaluation. Independent reporting frames it as closing much of the gap with leading coding models at a fraction of their price, with slower output speed as the main trade-off.
We flagged M2.7 while scoping model providers for One Inc's agentic stack — the price-to-performance is hard to ignore for long-horizon, tool-heavy work. The watch-out is throughput: it reasons well but emits tokens slowly, fitting background agents better than latency-sensitive ones.