MoeTraining-InfraOptimizationOpen-SourceGpu

Cursor's Mixture-of-Kittens Open-Source MoE Kernel Trains AI 2.37x Faster

Cursor released Mixture-of-Kittens (MoK), an optimized Mixture-of-Experts (MoE) megakernel for NVIDIA's NVL72 GPUs that solves computation and communication bottlenecks in sparse model training. The kernel achieves 2.37x faster training on MoE architectures, directly benefiting models like Cursor's Composer. This open-source contribution addresses a real bottleneck in scaling modern sparse models and demonstrates the performance gains possible with GPU-specific kernel optimization.

Why it matters

💻 Developer · If you're working with MoE models or planning to train sparse architectures, this kernel removes a major bottleneck. It's open-source and directly applicable—adopt it to cut training time by more than half and reduce iteration cycles.

📦 Product · Faster training means faster iteration on model improvements, cheaper experimentation, and quicker feature launches. This kernel cuts infrastructure costs while maintaining quality, directly improving your product velocity.

🎨 Design · Not directly relevant, but faster model training means your design team gets faster design-AI feature validation cycles and quicker prototyping of new capabilities.

📈 Business · This reduces AI training costs significantly—2.37x faster training means 2.37x fewer GPU-hours burned. For organizations training MoE models at scale, this translates to millions in hardware savings annually.

🤔 Just Curious · MoE models are the frontier of efficient scaling. This kernel shows how low-level GPU optimization is still a major research frontier—the biggest wins come from understanding hardware deeply, not just algorithms.

Try this: If you're training or fine-tuning sparse MoE models, integrate Mixture-of-Kittens into your training pipeline. Benchmark against your current training speed on NVL72—you'll likely see immediate 2-3x gains. Open-source means zero licensing friction.

Sources: Mixture-of-Kittens: our open-source MoE megakernel for NVL72s, Cursor's Mixture-of-Kittens Open-Source Kernel Trains AI 2.37x Faster