The Olmo-core 3 framework provides open, scalable infrastructure specifically for training large Mixture-of-Experts models. It optimizes memory usage and compute efficiency during the pre-training phase. This release allows researchers to build massive MoEs without proprietary software stacks.