Decoupled DiLoCo
AssessTechniques
A distributed low-communication training architecture that decouples compute islands and uses asynchronous data flow.
Why it's here
Placed in Assess: 1 article(s) of evidence from 1 source(s), led by research-stage coverage, with 0 in the last 30 days. Confidence 24%. Low accumulated evidence, so it defaults conservatively pending more signal.
Evidence (1)
- 8Google DeepMind·4/22/2026researchGoogle DeepMind unveils Decoupled DiLoCo for resilient distributed AI training
Google DeepMind introduced Decoupled DiLoCo, a distributed training architecture that splits large model training across asynchronous compute islands with lower bandwidth needs. The approach is designed to keep training running through hardware failures and data-center disruptions, while matching conventional training performance in Gemma 4 experiments.