Cursor Open-Sources MoK Kernel, Boosting MoE Training Throughput by 41%
rohanpaul_ai · x · 2026-08-06
AI coding tool Cursor has open-sourced Mixture-of-Kittens (MoK), its Mixture-of-Experts (MoE) training megakernel designed for NVL72 racks. It fuses all MoE communication and computation into a single, fully deterministic kernel, running up to 2.37x faster than the strongest public baselines.
Background & Technical Details:
- Training large models wastes massive GPU time moving tokens between chips. MoK rewrites the scheduling of these movements, squeezing 41% more work out of the exact same hardware.
- The replaced communication layer had become the largest single cost in training, swallowing more than half of total step time on some workloads.
- Earlier fixes attacked only the arithmetic (e.g., custom MXFP8 kernels), but once compute was no longer the bottleneck, data transfers became the primary bottleneck.
- The switch to GB300 NVL72 racks provided high-speed links across 72 chips but also introduced new systemic challenges.
Related event: Cursor Open-Sources MoK Kernel, Boosting NVL72 Training by Over 2x(8 posts)→
More from Infra
- Nous Research Launches Actual Local Inference Stack for Hermes Agent — NousResearch · 2026-08-06
- A Reddit user is seeking advice on whether to spend $2,800 now on a laptop with an AMD Ryzen AI Max+ 395 and 128GB of RAM, or hold off for Nvidia's upcoming RTX Spark. The dilemma centers on pricing predictions: buying now is a good deal if the Spark launches with a high price tag, but could lead to buyer's remorse if Nvidia shocks the market with an aggressively low price. — Dance-Till-Night1 · 2026-08-06
- Amazon CEO Expects $220B CapEx in 2026, Says Capacity Still Won't Meet AI Demand — Beth_Kindig · 2026-08-06
- Niche Infra Providers Face High Risks as AI Agents Become Cyber-Capable — tszzl · 2026-08-06
- TokenSpeed Scheduler Architecture: Decoupling C++ Control Plane from Python Execution Plane — zhyncs42 · 2026-08-06
- Musk: AI Compute Growing 10x Every Six Months, Traditional Data Centers Will Transition — r0ck3t23 · 2026-08-06