Cursor Open-Sources MoK: MoE Training Megakernel for NVL72 with 2x+ Forward Speedup
eliebakouch · x · 2026-08-05
Cursor has open-sourced Mixture-of-Kittens (MoK), a fully deterministic mixture-of-experts (MoE) training megakernel built from the ground up for NVL72 architecture.
- Total Fusion: Fuses all MoE computation and communication into a single, deterministic kernel, overlapping compute with inter-GPU networking at configurable granularity while fully eliminating CPU-GPU synchronization.
- Strong Performance: In standalone MoE layer benchmarks, forward pass runs up to 2.37x faster than the strongest public baselines.
- Precision: Supports BF16 and MXFP8 for both forward and backward passes, already powering production training of Cursor Composer.
Related event: Cursor Open-Sources MoE Training Kernel MoK for NVL72(2 posts)→
More from coding & agent
- You.com Search API Adds x402 Support for Autonomous AI Agent Payments — kleffew94 · 2026-08-05
- Goodfire AI Launches Silico for Autonomous Long-Horizon Experiments — mathildepapillo · 2026-08-05
- Opinion: RL Environments are all you need for training and evaluating agents — gregd_nlp · 2026-08-05
- llm-checker: inspect model files before loading to prevent crashes from malicious GGUF — tetsuoai · 2026-08-05
- Cloudflare Pay Enables AI Agents to Participate in Online Checkout — davidhoang · 2026-08-05
- Pokee Launches 28B Agentic Model with 10M Token Context, Runs on a Single RTX 4090 — Scobleizer · 2026-08-05