Cursor Open-Sources Mixture-of-Kittens, an MoE Megakernel for NVL72
nicolascraske · x · 2026-08-05
Cursor has open-sourced Mixture-of-Kittens (MoK), a production MoE (Mixture-of-Experts) training megakernel designed specifically for NVL72s.
- Core Optimization: Fuses all MoE communication and computation into a single, fully deterministic kernel, solving the communication bottleneck in large-scale training.
- Production Use: It currently powers the training of Composer, Cursor's agentic coding model, across tens of thousands of GPUs.
- Hardware Context: Rebuilt from first principles for the GB300 NVL72 multi-node rack architecture, leveraging its single NVLink domain across 72 GPUs for fine-grained overlap of compute and communication.
Related event: Cursor Open-Sources MoK Megakernel, Nearly Doubling NVL72 MoE Performance(5 posts)→
More from coding & agent
- Karpathy Tests Claude Opus: Writes 5500 Lines of Code to Render 3D Scenes — jon_barron · 2026-08-05
- Codex's Uncontrollable Urge to Check SKILL.md Files — dejavucoder · 2026-08-05
- Claude Fable 5 + Blender MCP: Describe Anything, AI Builds and Fixes It — jamestagg · 2026-08-05
- Open Source AI Hits a Wall in Long-Running Agentic Loops — bindureddy · 2026-08-05
- New ComfyUI Node: Generate PBR Materials and Sync to Blender/Unreal — Scared-Sandwich1283 · 2026-08-05
- Why Guardrails Fail: Rethinking Tool-Call Security in Coding Agents — eazyigz123 · 2026-08-05