Wire · opportunities
AI2 Ships Olmo-core 3, Claims 2.7x MoE Throughput vs Megatron
◆ Sectors
◆ Geography
◆ Source
◆ Verified
Fusion42 · 3 October 2026 · Fusion42 review
The Allen Institute for AI released Olmo-core 3, an open-source mixture-of-experts (MoE) training framework that achieves 2.7 times faster token processing throughput than Nvidia's Megatron-core on a 47 billion-parameter model and claims scalability beyond one trillion parameters.
This Wire brief sits within Fusion42's coverage of AI & ML, and 2 sources have reported it between 1 Oct 2026 and 3 Oct 2026.
◆ ◆ The Wire takeaway
You can now train large-scale mixture-of-experts models much faster with Olmo-core 3, opening a practical path to trillion-parameter models without typical efficiency losses. If your AI models bottleneck on training speed, switching to this open framework can cut your runtime significantly.
◆ Coverage
2 sources · first reported 1 Oct 2026 · latest 3 Oct 2026
◆ Related on Wire
◆ Topics