Issue 57Published October 2, 2026
Editor noteThis archived Ash AI Daily issue retains the delivered editorial briefing and final cards.
Story 1AI Daily
Ai2 opens Olmo-core 3 for large MoE training
Ai2 released Olmo-core 3, an open training framework for mixture-of-experts models. Its release says a 47B-parameter MoE reached 52,000 tokens per second per GPU on eight NVIDIA B300 GPUs, while tests also reached a 1.2T-parameter configuration across 512 GPUs. An MoE routes each token to a few specialized parts of a larger model; these are systems benchmarks, not quality evidence for a trained model.
- Ai2 released Olmo-core 3, an open training framework for mixture-of-experts models. Its release says a 47B-parameter MoE reached 52,000 tokens per second per GPU on eight NVIDIA B300 GPUs, while tests also reached a 1.2T-parameter configuration across 512 GPUs. An MoE routes each token to a few specialized parts of a larger model; these are systems benchmarks, not quality evidence for a trained model.
Why it matters
Open training infrastructure can make large sparse-model experiments more reproducible, but teams should distinguish performance engineering from a turnkey model deployment or independently replicated model-quality result.
Sources
Ai2 opens Olmo-core 3 for large MoE trainingAi2 / Hugging Face — Olmo-core 3 · Oct 2, 2026
