← Azimuth Daily
Capturing data…
Capturing data…
FRI, 2 OCT · 93 ITEMS
"AI2 releases Olmo-core 3, open training infrastructure for large MoE models" — Hugging Face Blog · AI & Technology
AI2 (the Allen Institute for AI) has released Olmo-core 3, an open-source framework and infrastructure for training large Mixture-of-Experts (MoE) language models. This release provides the training code, recipes, and tooling used to build AI2's upcoming generation of MoE-based open models, making the full stack publicly available for researchers and developers to replicate or extend.
Solid claim from a primary source: AI2's official Hugging Face blog post (noted as such) and confirmed by AI2's X announcement and community Reddit posts. The details about the infrastructure being open and focused on MoE training are directly from the organization's statement.
Genuinely new — the announcement was posted just hours ago as of the search results, with no recirculation year given.
We covered Sticky Routing (July 2026), a specific technique for memory-efficient MoE inference, which is directly relevant to the broader MoE infrastructure that Olmo-core 3 enables.