Google DeepMindLaunchDevelopersNovember 9, 2026

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

Read original source →developers.googleblog.com

MaxText v0.2.4 supports four context-parallel strategies (All-gather, A2A, P2P, USP) on Cloud TPU, hitting 70% MFU with Llama 3.1 8B at up to 4M tokens on TPU 5p. TPU 7 holds 50% MFU across sequence scales below 6M tokens.

1 source

More stories today

Open the live feed