MiniMax M2 Series: Production-Oriented Model Architecture

The MiniMax M2 family emerged as one of the most widely deployed open-weight language models in early 2026. The technical report details architectural decisions and optimizations that shaped its development, illustrating how production requirements now influence modern model design. The flagship implementation employs a sparse Mixture of Experts architecture, containing 229.9B total parameters while maintaining only 9.8B active parameters per token. The series encompasses distinct variants — M2, M2.5, and M2.7 — each representing different trade-offs between model capability and computational efficiency. A key distinguishing factor is how the model was engineered around real-world operational constraints: tool invocation latency, prefix caching mechanisms, executable environments, and iterative scaffolding. These were elevated to primary design considerations rather than afterthought optimizations, demonstrating industry maturation toward models built from inception to operate effectively within practical limitations.