MiniMax M2 is a large open-weight language model family developed by MiniMax, with the flagship variant built around a 62-layer decoder-only transformer. Independent technical-report notes describe it as carrying 229.9 billion total parameters while activating only 9.8 billion per token, an efficiency profile driven by a highly sparse mixture-of-experts layout that uses 256 fine-grained experts per MoE layer. The architecture retains full attention throughout all 62 layers, paired with grouped-query attention, QK-Norm, and partial RoPE, which together support stable long-context reasoning rather than relying on sliding-window or other sparse-attention shortcuts.
The model's design is closely tied to real-world software-engineering and agent workloads. Beyond its pre-training scale, the technical report documents training on executable software-engineering tasks and a reinforcement-learning setup that explicitly incorporates tool-call latency, tying architectural choices to how the model behaves inside agent loops. M2 also includes a multi-token prediction module that later siblings expand to multiple draft heads for speculative decoding. In practice, M2 is best understood as a coding- and agent-oriented model whose sparse-expert design, full attention, and tool-aware reinforcement learning make it well suited for code generation, multi-step developer assistance, and other structured tool-using workflows where long context and disciplined reasoning matter more than broad conversational breadth.
