Cursor Open-Sources MoK Kernel, Lifting NVL72 MoE Throughput by Up to 2.37x
Updated
Updated · cursor.com · Aug 4
Cursor Open-Sources MoK Kernel, Lifting NVL72 MoE Throughput by Up to 2.37x
3 articles · Updated · cursor.com · Aug 4
Summary
Cursor released Mixture-of-Kittens on GitHub, saying the production megakernel already runs Composer training across tens of thousands of GPUs and targets NVIDIA GB300 NVL72 systems.
Up to 2.37x faster MXFP8 forward throughput came from fusing MoE communication and computation into one deterministic kernel, after Cursor found inter-GPU communication—not compute—had become the main bottleneck.
On 512 GPUs across several NVL72 racks, MoK raised end-to-end training throughput to 1,070.2 tokens per second per GPU from 760.9, a 1.41x gain over Cursor's prior DeepEP-based setup.
MoK uses pull-based dispatch, push-based combine and ring token buffers to cut CPU-GPU synchronization and signaling overhead; Cursor said schedule generation takes under 3% of MoE runtime.
The open-source release targets DeepSeek-V3-style MoE layers used in models such as Qwen, GLM and Kimi, as Cursor seeks wider adoption and community contributions.