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

MiniMax released the M2 family of Mixture-of-Experts language models for agentic tasks.

The flagship M2 has 229.9B total parameters but activates only 9.8B per token, aiming for better efficiency.

Its training stack combines agent-generated data, the Forge reinforcement learning framework, and an M2.7 checkpoint that can help debug training.

MiniMax says the models perform strongly on coding, search, office tasks, and reasoning benchmarks.