Ai2 releases Olmo-core 3 with redesigned training for large AI models

Ai2 released Olmo-core 3, an open training framework for mixture-of-experts models, which activate selected specialized components for each input. The redesign keeps those components on GPUs and routes data to them, avoiding repeated gathering of model weights. In a preliminary test on eight NVIDIA B300 GPUs, Ai2 said a 47-billion-parameter model processed 52,000 tokens per second per GPU, compared with 19,400 using its earlier implementation. Researchers can use and adapt the stack for their own models. Ai2 also benchmarked a 1.2-trillion-parameter configuration across 512 GPUs. Those tests used random routing to measure system performance rather than trained-model quality; a separate 2.38-trillion-parameter configuration was a short capacity test.