AI2 Releases Olmo-core 3, Open Infrastructure for Training Large MoEs

In one line: The Allen Institute for AI (AI2) introduced Olmo-core 3, an open and scalable infrastructure for training large Mixture-of-Experts (MoE) models.

Key points

• Olmo-core 3 is presented as open training infrastructure focused on large MoE (Mixture-of-Experts) models.
• It leads with "scalable," suggesting a target of distributed training pipelines for large models.
• It appears to be a next-generation release of the training stack behind AI2's open Olmo model family.

Why it matters

MoE is a core architecture for improving compute efficiency in frontier models, yet most large-scale training code stays closed. Open training infrastructure gives smaller research teams and companies more room to reproduce and train MoE models themselves.

Read more

• Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs — Hugging Face Blog