All stories
RLX Tab Cluster Launches Community reports

RLX Tab Cluster Lets Browser Tabs Train Together

A new static demo turns each browser tab into a training participant and can link devices through WebRTC.

By AITrending1 min read
Sources (2)
Why now Hugging Face metadata dates the Space's creation to October 13. Its current README documents the browser-tab and cross-device training demo; the later modification time is not treated as a second launch.

A training cohort made of tabs

RLX Tab Cluster, a Hugging Face Space created on October 13, presents data-parallel training as an interactive browser demo. Its authors say each tab acts as a separate rank: a Web Worker compiles a backward graph, computes gradients over that tab’s own data shard, and averages them with the other participants after each round. Opening another tab joins it through BroadcastChannel. A room can bring in another device through a WebRTC mesh, with public relays handling signalling while gradients move peer to peer. The project page offers WebGPU, WebGL and CPU backend choices for individual ranks.

What the demo establishes

The Space also lets users choose among three gradient payload formats and shows the same model expressed in Rust, JavaScript and Python. Those are features described by the project, not independent proof of faster training, support for arbitrary models or useful scaling across devices. The README says cross-device rooms require HTTPS because WebRTC needs a secure context. As a static, self-contained demo, ClusterTab makes a distributed-training workflow visible in ordinary browser tabs; its real-world performance remains unverified here.