Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs
Allen AI releases Olmo-core 3: open-source training infrastructure for mixture-of-experts models at scale.
Why it matters
Olmo-core 3 lowers the barrier to training large MoE models by open-sourcing scalable infrastructure. Practitioners building or fine-tuning frontier-scale models gain a reproducible, transparent alternative to closed vendor stacks — critical for orgs evaluating in-house vs. managed training.
The key facts
6 to knowOlmo-core 3 is open-source training infrastructure for large mixture-of-experts models
Published by Allen AI (Allen Institute for AI)
Emphasis on scalability and openness; enables transparent, reproducible model training
No specific model size, benchmark results, or pricing disclosed in the title/announcement context
Targets practitioners training or fine-tuning MoE-based models; reduces vendor lock-in for training workflows
MoE architecture is frontier-relevant: conditional compute, efficiency, scaling approach
Go to the source
Hugging Face Bloghuggingface.co