MiniMax M2.1

MiniMax M2.1 is a 230B-parameter open-weight coding model with about 10B active parameters per inference, a 204,800-token context window, and strong polyglot coding results. In practice, I would treat it as a serious self-hostable coding model for agent workflows, not as MiniMax’s newest hosted model in 2026. What Is MiniMax M2.1? MiniMax M2.1 is a sparse mixture-of-experts coding and reasoning model from MiniMax. The important implementation detail is not only the headline size. MiniMax lists it as 230B total parameters with 10B activated per inference, which makes the serving profile very different from a dense 230B model. ...