Cursor Open-Sources Mixture-of-Kittens (MoK): A Deterministic MoE Training Megakernel for GB300 NVL72 Racks

Short summary

Cursor Research has open-sourced Mixture-of-Kittens (MoK), the MoE training megakernel powering its Composer models. MoK fuses all mixture-of-experts communication and computation into a single deterministic kernel, achieving up to 2.37x speedup over the best public baseline on GB300 NVL72 racks. It requires Blackwell SM100 or SM103 GPUs, limiting access to organizations with NVL72 infrastructure.

• •Cursor open-sourced MoK, a deterministic MoE training megakernel for Composer models
• •Achieves up to 2.37x speedup over best public baseline on GB300 NVL72 racks
• •Requires Blackwell SM100/SM103 GPUs, restricting access to NVL72-equipped organizations

Generated with AI, which can make mistakes.