Streaming DiLoCo Brings WAN-Based LLM Training to MaxText
Google Developers Blog says MaxText v0.2.5 introduces Streaming DiLoCo, which lets independently located learner groups train across datacenters while reducing cross-slice communication through periodic global updates and fine-grained parameter streaming. In tests using Qwen3-32B on two TPU v5p-32 slices, Google reported about 64–65 times less cross-slice traffic than data parallelism; the worked example says the 10 Gbit/s link required roughly 1.3 Gbit/s for Streaming DiLoCo. The feature is available through MaxText v0.2.5 on PyPI, its Docker images and the GitHub main branch.
Streaming DiLoCo gives AI infrastructure teams a released MaxText option for pooling TPU slices over constrained or geographically distributed networks without relying on one monolithic high-bandwidth cluster.
Distributed LLM Training Across Metros: Streaming DiLoCo in MaxText - Google Developers Blog
Distributed LLM Training Across Metros: Streaming DiLoCo in MaxText
NOV. 19, 2026
Mohit Khatwani Software Engineer
Chris Zuo Software Engineering Intern
Nova Fallen Software Engineer
#Streaming DiLoCo#MaxText#DiLoCo#Qwen3-32B#Google DeepMind#Google Research#Google Cloud#TPU v5p