Cursor open-sourced Mixture-of-Kittens (MoK), a deterministic mixture-of-experts training megakernel built for NVIDIA GB300 NVL72 racks, releasing the code on GitHub. MoK fuses all MoE dispatch/combine communication and expert FFN computation into a single kernel and is used in production to train Cursor's Composer coding model across tens of thousands of GPUs, where the MoE layer had accounted for more than half of end-to-end training time. Cursor reports up to 2.37x higher MXFP8 forward throughput than the fastest public baseline (comparisons included NCCL+PyTorch, DeepEP+PyTorch, DeepEP+TransformerEngine and HybridEP+Megatron) and a 1.41x increase in end-to-end tokens per second in its own stack. Design choices include pull-based forward dispatch with push-based combine, a device-side schedule kernel taking under 3% of MoE runtime, tunable minibatch sizes, and ring token buffers to remove CPU-GPU synchronization on the Grace CPUs.

• cursor.com2026-08-04