CURSOR OPEN-SOURCES MIXTURE-OF-KITTENS (MOK): A DETERMINISTIC MOE TRAINING MEGAKERNEL FOR GB300 NVL72 RACKS

Cursor Research has released an open‑source mixture‑of‑experts (MoE) training kernel called Mixture‑of‑Kittens (MoK). The new kernel, described by MarkTechPost, merges all MoE communication and computation into a single deterministic routine and can run up to 2.37 times faster than the best publicly available baseline on GB300 NVL72 racks.

According to the post, MoK powers the Composer models and achieves its speed boost by consolidating the entire MoE workflow into one kernel. It is designed to run on NVIDIA’s Blackwell SM100 or SM103 GPUs, meaning the technology is only accessible to users who have NVL72‑capable hardware. The deterministic nature of MoK removes variability in training performance, potentially simplifying debugging and deployment.

MoE training kernels are a key component of large language model training, and speed improvements can substantially reduce compute costs for organizations with the right hardware.

• Cursor Research open‑sources Mixture‑of‑Kittens, a deterministic MoE training megakernel.
• MoK integrates all communication and computation into one routine, delivering up to a 2.37× speed increase on GB300 NVL72 racks.
• The kernel requires Blackwell SM100 or SM103 GPUs, limiting its use to users with NVL72-capable infrastructure.

Read the original source for full detail available at MarkTechPost src