Cursor Open-Sources MoK: A Deterministic MoE Megakernel for NVL72s
StasBekman · x · 2026-08-05
Cursor announced the open-sourcing of Mixture-of-Kittens (MoK), a fully deterministic Mixture-of-Experts (MoE) training megakernel designed specifically for GB300 NVL72 architectures.
Key Optimizations & Context
- Bottleneck Fusion: As Cursor scaled Composer, its agentic coding model, the MoE layer became the major bottleneck, consuming over half of end-to-end training time. MoK fuses all MoE communication and computation into a single kernel to address this.
- Architecture Adaptation: NVL72 is a multi-node rack within a single NVLink domain, enabling fast, fine-grained overlap across all 72 GPUs. Additionally, the integrated Grace CPUs in GB300 are relatively slow, prompting a deep redesign of compute and communication.
- Production Proven: MoK currently powers Composer training across tens of thousands of GPUs.
Related event: Cursor Open-Sources MoK Megakernel, Nearly Doubling NVL72 MoE Performance(5 posts)→
More from Infra
- Best Local LLMs for Coding on a 128GB Mac? — Electronic_Back1502 · 2026-08-05
- Artificial Analysis launches Endpoint Accuracy Index: GLM-5.2, gpt-oss-120b, DeepSeek V4 Pro show big accuracy gaps across API providers — ArtificialAnlys · 2026-08-05
- Normal Computing Unveils CN101: A Digital Thermodynamic Computer for GenAI — ColesThermoAI · 2026-08-05
- a16z Co-leads Volta's Series A to Build AI Neocloud for Startups — a16z · 2026-08-05
- Nvidia's Strategic Shift: From Chip Sales to Cloud Revenue Sharing — GavinSBaker · 2026-08-05
- llama.cpp PR Caches Hot MoE Experts on GPU, Doubling Inference Speed on 8GB VRAM — BTA_Labs · 2026-08-05