MiniMax-M2

Overview

MiniMax-M2 is a Mixture-of-Experts model family introduced in a 2026 arXiv preprint at 229.9B total parameters with 9.8B activated (~4.3% activation ratio), paired with an agent-driven data pipeline and a Forge RL training framework that supports self-evolution and autonomous debugging. The activated-to-total ratio keeps inference cheap; pairing it with an agent-native RL recipe is the recipe to watch as sparse-activation models push toward frontier capability at sub-frontier serving cost.

Timeline

• 2026-05-27-AI-Digest — Paper “The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence” surfaces on HuggingFace Papers (arXiv:2605.26494, ▲11). MoE family at 229.9B total / 9.8B activated, with an agent-driven data pipeline and the Forge RL training framework supporting self-evolution and autonomous debugging.

Key Developments

1. Sparse-Activation + Agent-Native RL Recipe (May 27, 2026): The ~4.3% activation ratio keeps inference cheap; pairing it with an agent-driven data pipeline and self-evolving RL training is the practitioner-relevant combination as sparse-activation models push toward frontier capability at sub-frontier serving cost.

Related

See also: MOC - Open Source Models.