Cursor open-sources MoK after a claimed 41% training throughput gain
The Apache-licensed kernel fuses MoE communication and computation for training models on Nvidia GB300 NVL72 racks.
By Ryan Merket · Published
Primary source: X
Why it matters
Cursor is spending its funding on vertical model infrastructure, and MoK shows how application companies can cut training costs by optimizing communication across entire GPU racks.
Cursor (@cursor_ai) open-sourced Mixture-of-Kittens on August 4th, a GPU kernel that the AI coding company says increased end-to-end throughput for training its Composer models by 41% against its previous production stack.…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.