MiniMax-M2

MiniMax-M2 is positioned within MiniMax's M-series as an earlier Mixture-of-Experts language model that targets coding assistance, tool use, and agentic workflows. Its sparse architecture combines a large total parameter footprint of roughly 230 billion parameters with about 10 billion active parameters per token, a design that keeps per-token compute modest while preserving broad capability for software engineering tasks. The model was released on October 27, 2025, alongside MiniMax's "Agent" launch under the banner of simplicity, reflecting an intent to give developers a capable base for tool-augmented and multi-step problem solving rather than a general-purpose chat model.

Within the broader MiniMax lineup, MiniMax-M2 now sits as a legacy or comparison model, with newer siblings such as MiniMax M2.1, M2.5, M2.7, and MiniMax M3 having been released subsequently. Its emphasis on code generation, structured tool calling, and agent orchestration makes it a reasonable fit for teams maintaining existing pipelines on the M-series, building internal coding assistants, or benchmarking agent frameworks against an earlier-generation sparse model. Organizations evaluating it should weigh its maturity and tooling ecosystem against the more recent M-series variants when choosing a default model for new agentic applications.