The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence

MiniMax’s M2 series is a Mixture-of-Experts language-model line built around “mini activations, maximum real-world intelligence,” with the flagship M2 using 229.9B total parameters but only 9.8B activated per token, a 192K context window, 256 fine-grained experts, and pretraining on 29.2T tokens. The paper’s core novelty is not just scale but an agent-native stack: large verifiable data pipelines for software engineering, AppDev, terminal work, deep search, office tasks, and reasoning; Forge, a reinforcement-learning system that unifies white-box and black-box agents with windowed-FIFO scheduling, prefix-tree merging, and speculative decoding; and M2.7’s early self-evolution loop, where the model autonomously debugs training runs and edits its own scaffold. Across benchmarks, M2.7 remains competitive with Anthropic’s Claude Opus 4.6 and Sonnet 4.6, OpenAI’s GPT-5.4, and Google DeepMind’s Gemini 3.1 Pro despite the ~10B activated-parameter footprint, reaching 56.2 on SWE-bench Pro, 76.5 on SWE-bench Multilingual, 52.7 on Multi-SWE-bench, 57.0 on Terminal-Bench 2.0, 77.8 on BrowseComp, 50.0 on GDPval-AA, 46.3 on Toolathlon, 94.2 on AIME 2026, and 89.8 on GPQA-Diamond. The strongest within-series gains come from new data and RL pipelines, including a 13-point jump on NL2Repo, a 23.2-point gain on Finance Modeling Pro, and a 66.6% medal rate on MLE Bench Lite, tying Gemini 3.1 Pro while demonstrating autonomous ML-engineering and scaffold optimization.