Cursor open-sources Mixture-of-Kittens megakernel for MoE training

Cursor announced the open-source release of Mixture-of-Kittens (MoK), an MoE training megakernel optimized for NVIDIA NVL72 accelerators. The company claims MoK 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.

Cursor reported a 41% increase in overall tokens per second, which it says translates to billions of dollars in savings at scale. The announcement was made via a social media post on August 4, 2026.

The release follows a technical debate on the Latent Space Inference Engineering Masterclass about the viability of megakernels in production. Participants discussed the trade-offs between fused kernels and modular kernels, with some arguing that megakernels are impractical due to launch overhead and kernel complexity.

NVIDIA’s Rubin architecture is cited as addressing kernel overlap issues that previously motivated megakernels. Rubin enables finer-grained kernel coordination, including tile-level dependency triggers, which allow kernels to start as soon as data is available, reducing the need for fused kernels.

The discussion referenced physical constraints in GPU design that limit the effectiveness of megakernels, including the need for inter-GPU communication during nonlinear operations such as attention softmax.

Stuart Sul, a coauthor of the original megakernel work, now leads the team behind Mixture-of-Kittens, which is part of Dan Fu’s group at Cursor.