Ai2 releases Olmo-core 3, an open MoE training framework

The Allen Institute for AI has open-sourced Olmo-core 3 under the Apache 2.0 licence. The framework replaces fully sharded data parallelism with distributed data parallelism: a 47-billion-parameter MoE model reaches about 52,000 tokens per second per GPU versus 19,400 before (roughly 2.7x), while scaling the expert pool from 8 to 128 experts costs under 5% of throughput.

• Released 1 October 2026 under Apache 2.0, installable from PyPI as ai2-olmo-core
• 47B-parameter MoE: 52,000 tokens/s per GPU vs 19,400 on the old stack across 8 Nvidia B300 GPUs
• Growing the expert pool from 8 to 128 at ~3.2B active parameters costs under 5% of throughput
• Measured operating points span 12.9B parameters on 16 GPUs to 1.2T parameters on 512 GPUs