MiniMax M2.5 was introduced on February 12, 2026 as part of the MiniMax LLM family, sitting alongside the M2.7 and M3 models on the vendor's text-models page where it is positioned as state-of-the-art for coding and agent tasks and "Designed for Agent Universe." The model is built on a Mixture-of-Experts architecture with 230 billion total parameters but only about 10 billion active during any given inference pass, a sparse-activation design that the vendor and third-party reviewers credit with keeping serving costs low while still targeting frontier-class reasoning quality. By keeping most of the network idle per token, the MoE layout lets M2.5 aim at complex agent workloads without paying full dense-model compute on every request.

M2.5 is aimed squarely at developers building agentic coding assistants, tool-using workflows, and long-context automation. The vendor page frames it as a coding- and agent-focused release, and it ships within an ecosystem that lets teams run it through the MiniMax API or self-host it, giving engineering teams flexibility in how they deploy. FriendliAI's inference platform has added Anthropic Messages API compatibility for open models, which is the route through which this model can be served with minimal integration changes for teams already structured around that protocol. The combination of sparse activation, a large effective context, and agent-oriented positioning makes M2.5 a practical fit for code-generation pipelines, multi-step tool orchestration, and other developer workflows where both reasoning depth and economical per-token serving matter.