AReaL

AReaL is a reinforcement learning system from Tsinghua IIIS and Ant Group for training large reasoning and agentic models. Its core idea is fully asynchronous RL: generation and training are decoupled, and the v0.3 release reported a 2.77x speedup over synchronous systems with comparable or better results. A lighter AReaL-lite API exists for researchers who want algorithm-first prototyping.

AReaL 2.0 (July 2026) refactors it into independent training, inference, agent and weight-update services. A distinctive feature is online RL for black-box agents: you point an existing agent's base_url and api_key at AReaL's service and train it without changing agent code; the repo includes an OpenClaw example and Hermes and SWE training loops.

Inference runs on SGLang by default or vLLM, training can scale on Ray clusters, and Ascend NPU is supported on a separate branch. Setup assumes CUDA GPUs and a prebuilt flash-attn wheel, so it is research infrastructure rather than a library you drop into an app. License is Apache-2.0.