From the source
From the source
Open-source async RL system for LLM reasoning training
Ant Research released AReaL (Ant Reasoning RL), an open-source fully asynchronous reinforcement learning training system for large reasoning models.
The v0.3 release (boba²) achieves a 2.77x speedup over synchronous systems and includes SOTA coding models: a 14B model scoring 69.1 on LCB-v5.
The system supports multi-turn agentic RL training and scales from a single node to 1K GPUs.
From the source
AReaL (Ant Reasoning RL) is an open-source fully asynchronous reinforcement learning training system for large reasoning models developed at the RL Lab, Ant Research . Built upon the open-source project RealHF , we are fully committed to open-source by providing training details, data, and infrastructure required to reproduce results along with the model itself.
developer.ant-ling.com
