Details
- Cursor announces the open-sourcing of Mixture-of-Kittens (MoK), a Mixture-of-Experts training megakernel optimized for NVL72 GPU setups.
- MoK fuses all MoE communication and computation into a single, fully deterministic kernel, simplifying the stack and reducing complexity in distributed expert-parallel training.
- Cursor reports MoK runs up to 2.37x faster than the strongest public baselines for MoE training workloads, positioning it as a high-performance alternative to existing kernels.
- The company states that MoK now powers training across tens of thousands of GPUs in Cursor production environments, indicating maturity and large-scale operational use.
- In Cursor’s internal measurements, MoK increased end-to-end training throughput by 1.41x compared with its previous DeepEP-based stack, suggesting meaningful gains even over highly optimized expert-parallel communication libraries.
- Cursor frames the release as an effort to lower the barrier to AI research, enabling more labs and organizations to train large models efficiently on commodity or rented GPU clusters.
- The announcement includes links to the open-source MoK code repository and a technical write-up detailing how the megakernel was designed and implemented, providing both practical tools and architectural documentation for the community.
Impact
By open-sourcing a production-tested MoE training megakernel that outperforms existing public baselines, Cursor strengthens the ecosystem of large-scale training tooling and narrows performance gaps between well-funded AI labs and smaller research groups. The reported gains over DeepEP-based setups may push other frameworks and cloud providers to revisit their MoE kernels and communication patterns, accelerating optimization work around expert-parallel training on NVL72-class hardware.
