Observed facts: SiliconANGLE reports Ai2 says Olmo-core 3 can expand its expert pool from eight to 128 while selecting four experts per token, and can scale to models above one trillion parameters. Observed facts: Ai2’s reported benchmark for a 47-billion-parameter model is 52,000 tokens per second on Nvidia B300 GPUs, versus about 19,400 for Nvidia Megatron-core.
Mechanism: The account describes expert parallelism across GPUs, layer splitting across GPU groups, a distributed optimizer that avoids full copies of optimizer state on every GPU, and MXFP8 support. Together, these techniques target memory overhead and data movement as the expert pool grows, while maintaining sparse activation. Why now: Access to the code and related systems is now reported for developers and the open-source community, making the release an implementable training-stack option rather than only a design proposal.
First implication: For teams building large MoEs, the framework offers a concrete alternative to treating dense-model tooling as the default; throughput and fit will depend on the cluster and workload. Evidence boundary: The figures and scale claims are attributed to Ai2 through SiliconANGLE; the supplied source does not establish independently reproduced results or performance across other hardware, model sizes, or workloads.
First watch condition: Look for independently reproducible benchmarks and evidence that the memory and throughput gains persist on varied GPU configurations. This starts a distinct analytical record because no prior New Runtime record was supplied for this release.