The Allen Institute for AI released Olmo-core 3 to streamline the training of mixture-of-experts models. This framework optimizes how large language models distribute workloads across specialized parameters. It reduces the computational overhead typically required for MoE architectures.