Reproducing Olmo 3 7B Pre-training in MaxText: case study of large scale training on TPUs
First reported · 21d ago · 1 source
Crest summary — written by AI from the 1 source below; every figure is checked against them · model deepseek-chat
Summary
Google's MaxText team reproduced Ai2's Olmo 3 7B language model from scratch on Google Cloud TPUs using JAX/XLA, matching the original training run.
Key facts
- The reproduction used Google Cloud TPUs with the JAX/XLA stack.
- The model reproduced was Ai2's Olmo 3 7B language model.
- The work was published on the Google Developers Blog on 2026-10-03.
- Training was done from scratch and matched the original run.
Why it matters
Teams evaluating TPUs for large-scale model training now have a documented case of reproducing a 7B open model end to end, which can inform hardware and framework choices.
Companies
Technologies
- JAXlibrary
- XLAframework
- MaxTextframework
- Olmo 3 7Bmodel
- TPUsproduct
Related on Crest
Matched by rule, not by AI: a repository linked in the coverage, or a technology or company in this event that Crest tracks.
Repositories
- google/jaxoptHardware accelerated, batchable and differentiable optimizers in JAX.A named company’s project
- jax-ml/jaxComposable transformations of Python+NumPy programs: differentiate, vectorize, JIT to GPU/TPU, and moreMatches a named technology
- openxla/xlaA machine learning compiler for GPUs, CPUs, and ML acceleratorsMatches a named technology
Original sources (1)
The report this event was built from. Crest of Tech keeps no article bodies: the full text is at each source.
- Google Developers BlogOfficial · PrimaryReproducing Olmo 3 7B Pre-training in MaxText: case study of large scale training on TPUsPublish time not givenRead at Google Developers Blog