AReaL

AReaL is a large-scale asynchronous reinforcement learning (RL) system for training reasoning and agentic models. It was originally developed by researchers and engineers from Tsinghua University's IIIS and the AReaL team at Ant Group, and its fully asynchronous training paradigm decouples generation from training to improve efficiency and scalability.

It is aimed at ML engineers and researchers who want to train or fine-tune models for agent tasks with RL, such as math, coding, search and customer-service agents, where the project reports state-of-the-art results. A notable feature is online RL for black-box agent applications: you can train against an existing agent by replacing its model `base_url` with AReaL's endpoint. AReaL 2.0, released in July 2026, refactored the system into independent training, inference, agent and weight-update services and added an online RL loop for the Hermes agent.

Getting started: AReaL targets Linux machines with NVIDIA GPUs. Install uv, install the matching flash-attention wheel, run `uv sync --extra cuda` (SGLang is the default inference backend), and follow the documentation's tutorials for agentic and online RL.