TL;DR
A technical report on the MiniMax-M2 family, pairing a 229.9B-parameter sparse MoE with agent-generated, verifiable trajectories and Forge, a scalable reinforcement-learning system for long-horizon agents.
Why it matters
M2 treats the model, executable environment, trajectory data, artifact-aligned reward, and RL infrastructure as one scaling system. It is a concrete account of how a highly sparse model can be trained for coding and knowledge-work agents without reducing the problem to model size alone.
Key findings
- 01
The training pipeline produces large-scale trajectories for agentic coding and agentic cowork inside executable workspaces, with rewards aligned to the resulting artifacts.
- 02
Forge combines long-horizon agent RL with windowed-FIFO scheduling, prefix-tree merging, inference optimization, and separation between training, inference, and agent execution.
- 03
The authors describe M2.7 as an early self-evolution experiment that can debug training runs and modify its own scaffold; its reported benchmark gains should still be read as vendor-reported results.
Scaling dimensions
Models, methods & benchmarks
- Models
- MiniMax-M2, MiniMax-M2.7
- Algorithms
- Forge, Agent-native reinforcement learning, Windowed-FIFO scheduling, Prefix-tree merging
Topics