The Year in AI Papers
The year/MiniMax

Paper 2605.26494

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

Aili Chen, Aonian Li, Baichuan Zhou, Bangwei Gong, Binyang Jiang, Boji Dan, Changhao Zhang, Changqing Yu, Chao Wang, Cheng Ma, Cheng Zhong, Cheng Zhu, et al.

Published
May 2026
Research lab
MiniMax
Citations
23
GitHub
Not linked

01 In brief

Summary

The MiniMax-M2 series is a family of Mixture-of-Experts language models designed to achieve frontier performance with minimal activated parameters.

The flagship M2 has 229.9B total parameters but only 9.8B activated per token, using a 62-layer decoder-only Transformer with 256 fine-grained experts, sigmoid gating, full multi-head attention with GQA, a 192K-token context window, and a Multi-Token Prediction module for speculative decoding.

Pre-trained on 29.2T tokens, the series' real-world capability comes from an agent-native post-training pipeline with three components: agent-driven data pipelines for coding, cowork, and reasoning tasks with verifiable rewards; Forge, an RL system supporting white-box and black-box agents with windowed-FIFO scheduling and prefix-tree merging; and self-evolution in M2.7, which autonomously debugs training runs and modifies its scaffold.

Results show M2.7 is competitive with larger closed-weight models, scoring 56.2 on SWE-bench Pro, 52.7 on Multi-SWE-bench, 77.8 on BrowseComp, 50.0 on GDPval-AA, 94.2 on AIME 2026, and 89.8 on GPQA-Diamond, with only ~10B activated parameters.

02 From the paper

Abstract

We introduce the MiniMax-M2 series, a family of Mixture-of-Experts language models built around the principle that mini activations can unleash maximum real-world intelligence. The flagship M2 contains 229.9B total parameters with only 9.8B activated per token. Designed end-to-end for agentic deployment, the M2 series rests on three components: (i) agent-driven data pipelines producing large-scale, verifiable trajectories across agentic coding and agentic cowork, each grounded in an executable workspace and an artifact-aligned reward; (ii) Forge, a scalable agent-native RL system that adapts to long-horizon agent trajectories, paired with windowed-FIFO scheduling, prefix-tree merging, inference optimization, and a clean training-inference-agent decoupling that supports both white-box and black-box agents; (iii) the latest M2.7 checkpoint takes an early step toward self-evolution -- autonomously debugging training runs and modifying its own scaffold. Across M2 through M2.7, this combination translates a mini-activation footprint into frontier-tier performance on agentic coding, deep search, office-task, and reasoning benchmarks.