Cursor Open-Sources Mixture-of-Kittens, a Deterministic MoE Training Kernel

Cursor Research released Mixture-of-Kittens (MoK), an open-source mixture-of-experts training megakernel that powers its Composer models. The kernel fuses all MoE communication and computation into a single deterministic operation, achieving up to 2.37x higher throughput than existing public alternatives. Available under Apache 2.0 license on GitHub, MoK requires NVIDIA Blackwell GPUs (GB200 or GB300 NVL72 racks) and is already deployed across tens of thousands of GPUs for Composer model training. The technology targets high-value applications including deep learning model pretraining, reinforcement learning post-training, and ablation studies.