// AI news · 2026-11-09 · Google AI Changelog
Google enables 14M-token long-context training on TPUs via MaxText
Google introduces long-context training up to 14M tokens on Cloud TPUs via MaxText, leveraging context parallelism and Tokamax Splash Attention.
Published 2026-11-09 08:00 (Beijing time)Source Google AI ChangelogAuto-summarised, fact-checked against source
01Summary
The solution enables compute-communication overlap with stable per-chip memory usage. Llama 3.1 8B achieves 70% MFU on TPU 5p at 4M tokens, and 50% MFU up to 14M tokens on TPU 7.
Summarised from the original article by an open-source model, then checked automatically: every number and English name must appear in the source. The original article is authoritative.
02Key points
- MaxText enables 14M-token training on Cloud TPUs with context parallelism.
- Tokamax Splash Attention and 3D ICI enable compute-communication overlap and stable memory per chip.
- Llama 3.1 8B reaches 70% MFU on TPU 5p at 4M tokens.
- Four context strategies (All-gather, A2A, P2P, hybrid) support load balancing and sequence packing.
03Source
Announcing Long-Context Training on TPUs with MaxText Context Parallelism
Google AI Changelog · published 2026-11-09 08:00 Beijing time
05Related news
- Same source · Build Geospatial Apps with Google Maps in AI Studio2026-11-09 · Google AI Changelog
- Same source · Google launches on-device ASR and TTS in Edge Gallery2026-11-10 · Google AI Changelog
- Same topic · Google Releases TPU Sync for Fast Model State Transfer2026-11-09 · Google Cloud
- Same topic · Google Launches Cloud Pathways ML Runtime at GA2026-11-09 · Google Cloud
- Same day · Meta Muse Ember available on Vercel AI Gateway2026-11-09 · Vercel Changelog
From AIMap AI news: collected from 1,100+ official and media sources. Summaries only restate the original article; follow the link for the full story. All stories of 2026-11-09 · Archive by date · All AI news · Today