Olmo-core 3: Ai2 open MoE training stack (not a chat-model GA)

Ai2 released Olmo-core 3 on 2026-10-01: a significant upgrade to its framework for developing large language models, featuring a redesigned open mixture-of-experts (MoE) training system aimed at scaling MoE training into the trillion-parameter range (HF blog).

This is training infrastructure, not a new public Olmo chat or instruct model. Ai2 frames Olmo-core 3 as the foundation for a next-generation MoE Olmo it aims to make its most capable yet—with no ship date or public weights announced in this post (HF blog).

What changed in the stack

Earlier Olmo-core MoE training used FSDP (gather and reshard weights per small batch). Olmo-core 3 switches to a DDP-based path that keeps experts resident on GPUs and routes data to them (HF blog).

Ai2 also describes expert parallelism, pipeline parallelism, a distributed optimizer, rowwise expert parallelism, GPU-resident routing, grouped GEMM, and MXFP8 support—as systems features of the open stack (HF blog).

Throughput and scale (Ai2 benches)

All figures below are Ai2-stated vendor benches on stated NVIDIA B300 configs—not independent newsroom measurements (HF blog):

Claim	As Ai2 states
Prior MoE path vs new stack	Preliminary test, 47B MoE on 8× B300: 52,000 vs 19,400 tok/s/GPU ≈ ~2.7×
MXFP8 vs BF16	Controlled bench on 4× B300: throughput ~21% higher; peak active memory 103 → 95 GiB
Trillion-scale demo	1.2T total params, 58.36B active/token, 512 GPUs; peak 858 TFLOP/s/GPU; random routing to measure system performance, not trained-model quality
DeepEP v2 experiment	Configuration with 2.38T total params—a short-capacity test, not a full training run

License and where to get it

Code on GitHub allenai/OLMo-core is Apache-2.0. This brief covers the open training stack with attribution; it does not redistribute code or grant licenses. Ai2 also links a technical report and interactive walkthrough from the blog—deep-link those rather than rehosting figures (HF blog, GitHub).

Who should care

Labs and researchers who want an open MoE training stack with Ai2’s DDP expert-resident redesign should start at the Olmo-core 3 post and GitHub repo—and treat every TFLOP/throughput number as Ai2’s preliminary or controlled bench, not a chat-model launch.