Cursor Open-Sources Mixture-of-Kittens MoE Megakernel
Fused kernel combines communication and computation for MoE training on NVL72 hardware.
Cursor announced the open-sourcing of Mixture-of-Kittens, its megakernel for mixture-of-experts training on NVL72 hardware. The kernel fuses all communication and computation steps into one fully deterministic unit. Cursor states that the code now powers its training runs across large GPU fleets and improves end-to-end throughput over the prior DeepEP-based stack. Independent posts note the technique reduces time spent moving tokens between chips. The company shares the code publicly, with replies from engineers and founders welcoming the contribution to the open ecosystem.
We're open-sourcing Mixture-of-Kittens (MoK), our MoE training megakernel for NVL72s. It fuses all Mixture-of-Experts communication and computation into a single, fully deterministic kernel, and runs up to 2.37x faster than the strongest public baselines.

