Cursor Open-Sources Mixture-of-Kittens (MoK): A Deterministic MoE Training Megakernel for GB300 NVL72 Racks
Cursor open-sources MoK, a 2.37x faster MoE training kernel—but only for teams with NVL72 racks.
Why it matters
A significant optimization in frontier model training (MoE efficiency gains), but with a harsh availability gate: the kernel requires Blackwell-class GPUs that most practitioners cannot access. Relevant to labs training large models and to the broader frontier-lab race on training efficiency; less immediately actionable for most practitioners.
The key facts
6 to knowCursor Research open-sourced Mixture-of-Kittens (MoK)
MoE training megakernel achieves 2.37x speedup vs. public baseline
Fuses MoE communication and computation into single deterministic kernel
Requires Blackwell SM100 or SM103 GPUs (NVL72 racks only)
Powers Cursor's Composer models
Published August 4, 2026
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Cursor Research has open-sourced Mixture-of-Kittens (MoK), the MoE training megakernel behind its Composer models. MoK fuses all mixture-of-experts communication and computation into a single deterministic kernel, and runs up to 2.37x faster than the strongest public baseline on GB300 NVL72 racks.…