The MiniMax-M2 Series: Technical Report

"Mini Activations Unleashing Max Real-World Intelligence." Comprehensive technical report covering the full M2 / M2.1 / M2.5 / M2.7 series — an agent-native MoE family at 229.9B total / 9.8B active. Three load-bearing contributions documented end-to-end:

• Agent-driven data pipelines producing agentic coding and cowork trajectories at scale
• Forge, a reinforcement-learning system built specifically for long-horizon agentic tasks
• Self-evolution at the M2.7 checkpoint — the model autonomously debugs and improves prior outputs in a closed loop

Reports frontier-level results across agentic coding, deep search, office-task, and reasoning benchmarks. 35 pages, 10 figures, 4 tables, with ~205 listed contributors. The canonical reference for the whole M2 line.