AReaL v0.5.0 Opens the Async RL Engine Behind Ring-1T Post-Training

Training an agent with reinforcement learning usually means rewriting its tool-calling code against whatever interface the RL framework expects. Ant Open Source releases AReaL v0.5.0, the asynchronous RL engine the team says trained InclusionAI's Ring-1T trillion-parameter model, and describes it exposing an OpenAI-compatible API proxy so an existing agent can be plugged into RL training without custom rollout plumbing. The release follows AState, the third piece of the team's open-source ASystem effort, with AReaL as the fourth.

The company also describes moving from an SPMD execution model, where every process runs the same program in lockstep, to a single-controller architecture, which it says removes long-tail latency and data imbalance that can appear when GPUs fall out of sync during large rollout batches. AReaL builds on the team's earlier ReaLHF framework according to the thread, though the specific architectural differences between the two are not detailed in the post and are not verified here beyond what is stated.