Ai2 open-sources Olmo-core 3 training framework for more efficient MoE training, with about 2.7x per-GPU throughput on a 47B model (report)

Verified (fact)

According to AI Times, the Allen Institute for AI (Ai2) has released Olmo-core 3, an LLM training framework, as open source. Ai2 redesigned its existing training system, and per-GPU throughput on a 47-billion-parameter (47B) MoE model reportedly rose by about 2.7x. The report said the framework also targets models with trillions of parameters. The summary does not give the exact release date or the baseline for the comparison (whether a previous version or another framework).

Needs fact-check (claim)

Ai2 claims: about 2.7x higher per-GPU throughput on a 47B-parameter model (comparison conditions not given in the report summary).

Why it matters (interpretation)

MoE has become a mainstream architecture for training large models. If an open framework can substantially raise training throughput, labs and companies without abundant GPUs will find it easier to train large models themselves. However, what the 2.7x figure is measured against needs to be checked in the original source.

Sources

• Press Ai2, MoE 학습 효율 2.7배 높인 오픈소스 프레임워크 '올모 코어 3' 공개 AI Times (Korea)