MiniMax details Forge, its agent RL framework
MiniMax has published the design of Forge, the in-house reinforcement learning system behind its M2.5 model, describing how it trained across more than 100,000 agent scaffolds at 200k context.
MiniMax described its internal agent RL framework, Forge, in a technical post. The company says Forge is what made the capabilities of its previously announced MiniMax M2.5 model possible.
According to MiniMax, Forge worked with more than 100,000 distinct real-world agent scaffolds and environments while M2.5 was being built. It handled context lengths up to 200k tokens and processed on the order of millions of samples per day, and the company says reward stayed convergent throughout. The training paired Forge with MiniMax's CISPO algorithm and a composite reward design.
The post presents agent RL as a trade-off between system throughput, training stability and agent flexibility. MiniMax's answer is a middleware layer that separates agent logic from training infrastructure. A Gateway server exposes standard completion endpoints, so the agent never sees which model sits behind them. A distributed Data Pool collects trajectories asynchronously. This setup lets Forge train both white-box and black-box scaffolds, which the authors say is hard with token-in/token-out designs that couple agents tightly to the trainer.
The post also covers scheduling and redundancy. Strict FIFO scheduling leaves the cluster idle behind slow rollouts. Greedy first-finished scheduling skews the training distribution toward short, easy tasks. MiniMax's fix is a windowed-FIFO scheduler. The company also reports a 40x training speedup from merging shared prompt prefixes into a tree.
All of these figures come from MiniMax. The company has not released code for Forge, and no independent benchmarks of the system are available.