Decoupled DiLoCo
Decoupled DiLoCo trains LLMs across separate asynchronous compute islands, tolerating hardware failures and cutting bandwidth between data centers by orders of magnitude.
Builds on Pathways (asynchronous data flow) and DiLoCo (low-communication training). It self-heals. In chaos-engineering tests with Gemma 4, training continued after whole learner units failed, and it matched benchmark quality while keeping more goodput.
- Date
- Thursday, 23 April 2026
- Lab
- Google DeepMind
- Kind
- paper
- Access
- research preview
Some plotted results are simulations per the blog captions. Page dated 2026-04-23 (RSS lists Apr 22).
Sources
This record was checked against its sources on 6 October 2026. How we check