Google implements Streaming DiLoCo in MaxText for cross-datacenter LLM training
11/21/2026 — 11/21, 15:17·1 sources (all)·1 reports (all)
Key takeaways
- Google DeepMind and Google Research have implemented Streaming DiLoCo in the open-source JAX training framework MaxText. By streaming parame…
Background & analysis
On November 21, 2026, at 15:10, Google Developers Blog reported that teams from Google DeepMind and Google Research have implemented Streaming DiLoCo in MaxText, the open-source JAX training framework, to support LLM training across data centers.
According to the post, training state-of-the-art LLMs has traditionally depended on a single, monolithic accelerator cluster, where every device sits in one location and is tightly coupled through high-bandwidth links that keep gradients in constant sync. The authors note that physical limits and rising cost pressures are making it increasingly hard to scale a single slice any further. Streaming DiLoCo addresses that constraint by streaming parameter fragments instead of synchronizing the full model at every training step. This cuts the volume of communication across TPU slices by more than 60× and removes network stalls on the WAN, the post states. Google says the approach allows training to be scaled across multiple data centers without degrading model convergence or model quality.
At this point, the only material available is the announcement itself and the explanation published by Google. There is no follow-up release, benchmark detail, or independent verification in the source material to confirm how the method performs in practice.
AI-generated from 1 reports · updated 19 hours ago
- Heat index
- 57
- All sources
- 1
- All reports
- 1
- First seen
- 20 hours ago
Reports on this story headlines open the original
Google DeepMind and Google Research have implemented Streaming DiLoCo in MaxText, Google's open-source JAX-based LLM training framework. Instead of synchronizing full models at every step, it streams parameter fragments, cutting cross-slice communication volume by more than 60x and removing WAN stalls without hurting convergence or model quality, according to the company's blog post.
Google Developers BlogFirst-partyAI score 62
Other stories people are talking about
- Heat index 1143RisingAnthropic Reaches $1.6 Trillion Market Value in First Trading Week After IPO21 sources
- Heat index 1011Anthropic Mathematician Says Claude Formalized the Generalized Riemann Hypothesis21 sources
- Heat index 622Anthropic IPO total raise reaches $93.5 billion after greenshoe exercise20 sources
- Heat index 596RisingThinking Machines Lab Open-Sources Inkling 2 Mini with 32GB Model Card Recommendation10 sources
- Heat index 490RisingSEC probes Nvidia accounting practices as shares post seven-week slide9 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion

