Definition
Plain language
MiniMax's custom training system for teaching AI agents through long, varied real-world trajectories.
As stated in the literature
MiniMax's RL infrastructure for agent training, with decoupled model, agent, and trainer modules, windowed FIFO scheduling, and prefix tree merging for shared-context efficiency.
Why it matters: Training agents on long, varied trajectories at scale is bottlenecked by infrastructure, and clever scheduling makes the difference between feasible and impossible.
For example, when a thousand agent runs all start from the same long system prompt, Forge stores that shared prefix once and stitches in only the diverging tails to save memory.
Heard on the show
“They built a custom RL system called Forge — and this is where the most concrete engineering shows up.”Episode 090 — How MiniMax-M2 Bets That Sparsity Plus Verifiable Rewards Can Match Frontier Agents