Cursor's MoK Optimizes GPU Pipeline, Boosts MoE Training by 41%

According to Beating Monitor, Cursor has open-sourced Mixture-of-Kittens (MoK) to accelerate MoE large model training. MoK combines previously separate GPU data transfers and computations into a single kernel—a low-level program running directly on the GPU. MoE splits models into numerous “experts,” activating only a subset at a time. Since these experts are distributed across different GPUs, data must be frequently moved, sometimes consuming over half of the training time. MoK enables GPUs to transfer data and perform computations simultaneously, reducing idle waiting. In real-world training using 512 GB300 GPUs, MoK improved overall throughput by 41%. When tested independently on MoE layers, it achieved up to 2.37x faster performance than public alternatives. MoK has already been deployed in Cursor’s Composer training across tens of thousands of GPUs and is now open-sourced under Apache 2.0. Currently, it supports only NVIDIA Blackwell GPUs and is primarily designed for institutions with GB200, GB300, or NVL72 clusters.