Google engineers have reproduced the Allen Institute for AI's Olmo 3 7B language model on Google Cloud TPUs using the MaxText framework.

The reproduction covered both stage-1 pre-training and stage-2 mid-training anneal. Ran Ran, a Staff Software Engineer at Google Cloud, led the team that performed the work.

This run appears to confirm the Olmo 3 training recipe can be reimplemented outside the original lab on standard cloud accelerator hardware.

An optional feature will be added to the Google Messages app to analyze suspicious messages using the Gemini artificial intelligence model.