MiniMax M2 and Production-Oriented Model Design

This article summarizes key insights from the MiniMax M2 technical report, focusing on production-oriented model design. It covers the sparse MoE architecture (229.9B total, 9.8B active parameters), the decision to use full attention over hybrid sliding-window variants due to production tradeoffs, and challenges with linear/sparse attention for coding agents. Fine-grained MoE with 128 experts and top-8 routing shows significant improvements in MATH and HumanEval benchmarks. The report details agent training pipelines using GitHub pull requests and Docker environments, interleaved thinking for context management, speed rewards in RL, and self-evolution where M2.7 handles daily RL iteration. The broader takeaway is that production constraints like prefix caching and tool latency are now first-class inputs in model design.