Ant Group Open-Sources Awex Framework for Second-Level TB-Scale RL Parameter Swapping
Want to read in a language you're more familiar with?
Ant Group open-sources Awex, a framework enabling second-level TB-scale parameter exchanges for trillion-parameter RL models, slashing training latency and boosting efficiency.
On November 20, 2025, Ant Group announced the open-sourcing of Awex, a trillion-parameter reinforcement learning weight-swapping framework that completes TB-scale parameter exchanges in seconds, significantly reducing RL model training latency.
Awex leads in performance: Under RDMA on a thousand-card cluster, it syncs 1TB model weights in 6 seconds, with full trillion-parameter synchronization also in just 6 seconds. It supports NCCL, shared memory, and other modes, compatible with multi-model architectures and heterogeneous deployments, minimizing overhead via zero-redundancy transmission and in-place updates.
As a core component of Ant Group's ASystem RL system (powering the Bailian trillion-parameter model training), Awex integrates with Megatron and SGLang engines. Ant plans to open-source more ASystem RL components to bolster the RL open ecosystem.
Source: IT Home
