AI2's Olmo-core 3: Open Training Infrastructure for Large MoE Models

It is a new release of the open-source training infrastructure developed by the Allen Institute for AI (AI2), designed to support training of large Mixture-of-Experts (MoE) models with better scalability and efficiency.

Olmo-core 3 emphasizes full openness: teams can inspect the code, reproduce experiments, and modify the stack. Commercial training pipelines typically remain closed and cannot be audited or adapted.

Yes. Because the stack is open source and published via Hugging Face, research groups, government labs, and enterprises can experiment locally, provided they have access to high-performance compute.

The source does not include independent benchmark results or detailed performance comparisons, so claims about superiority or cost savings should be treated cautiously until third-party measurements appear.