Google MaxText adds long-context training up to 14M tokens on TPUs

Pushes open training stack to multi-million-token contexts, easing long-context model work on TPUs.

Google Cloud's open-source MaxText framework now supports long-context training up to 14 million tokens on Cloud TPUs via four context-parallel strategies (All-gather, A2A, P2P, USP), powered by the Tokamax Splash Attention kernel.

Domain

general

Read it at Google Developers →

Coverage: 1 sources, in order

1. 09 Nov 21:00 Google Developers Read the article →

← Front page