Ai2 releases Olmo-core 3, an open MoE training stack designed for trillion-parameter scale

What changed

Ai2 released Olmo-core 3, a redesigned open training stack for large mixture-of-experts models with expert and pipeline parallelism, distributed optimization, GPU-resident routing and MXFP8 support. Ai2 says the stack has been benchmarked at 1.2 trillion parameters across 512 B300 GPUs, and a preliminary eight-B300 47B MoE test reached about 2.7× the throughput of its earlier FSDP implementation. The company explicitly frames these as systems-performance tests rather than model-quality results.