Ai2 releases open-source MoE training framework Olmo-core 3

On October 1, 2026, Ai2 (Allen Institute for AI) released the open-source training framework Olmo-core 3 via the Hugging Face Blog. The framework targets large-scale mixture-of-experts (MoE) model training, redesigns the open-source MoE training system, and claims to support scaling up to the trillion-parameter level. On the technical approach, Olmo-core 3 abandoned the previous FSDP (Fully Sharded Data Parallel) scheme in favor of an implementation based on DDP (Distributed Data Parallel). On performance, Ai2 said a 47B-parameter MoE model achieved a throughput of 52 on 8 NVIDIA B300 GPUs. The framework is now publicly available; the report does not mention earlier version history or other follow-up plans.

AI overview of reporting · Updated 17 days ago