Introduction to Mixture‑of‑Kittens
The field of machine learning is evolving rapidly, and Mixture‑of‑Experts (MoE) models represent a major breakthrough for managing extremely deep networks while maintaining efficiency. Cursor Research has recently released Mixture‑of‑Kittens (MoK), an open‑source megakernel that consolidates all communication and computation stages of an MoE into a single deterministic kernel.
This repository, hosted on GitHub under the Apache 2.0 license, promises up to 2.37× higher throughput than existing public solutions. It is already being used to train Composer models at scales of tens of thousands of GPUs, demonstrating its robustness and industrial potential.
Technical Architecture of the Megakernel
MoK merges each phase—from expert distribution to post‑forward aggregation—into a continuous stream. This approach eliminates frequent synchronizations, reducing wait times between nodes and optimizing inter‑GPU bandwidth usage.
Symmetric Memory Management
The implementation relies on PyTorch’s symmetric memory feature, enabling efficient buffer sharing across GPUs. This technique ensures deterministic coherence while minimizing unnecessary copies.
CUDA 13 and Blackwell GPU Support
To fully leverage MoK, you need NVIDIA Blackwell SM100 or SM103 GPUs (GB200 NVL72 / GB300 NVL72). These GPUs provide the bandwidth and compute capacity required to sustain the megakernel’s high throughput.
Performance and Benchmarks
Cursor’s published tests show an average 1.75× increase on standard MoE pre‑training tasks, with a peak of 2.37× on DeepSeek‑V3‑style models. Gains are attributed to reduced synchronization points and optimized data flow.
- GPU Throughput: +1.75×
- Inter‑GPU Latency: –40 %
- Memory Efficiency: +20 % effective usage
Practical Deployment and Constraints
MoK is deployable but requires a high‑end environment. In addition to Blackwell GPUs, it demands Python 3.12+, PyTorch 2.10+ and CUDA 13.0+. These requirements limit adoption to organizations already equipped with NVL72 infrastructure or able to lease them.
Software Ecosystem
The integration is done via a lightweight Python package that exposes the megakernel API. Users can replace their traditional MoE routines without modifying model code.
Ideal Use Scenarios
Research labs, venture‑backed startups, and national supercomputing centers are the primary beneficiaries. Teams with fewer than 8 GPUs will not find this technology worthwhile due to the minimum parallelism threshold required.
Use Case: Composer Models and RL
“Mixture‑of‑Kittens enabled us to reduce training time for our Composer model by 35 % while ensuring reproducibility of results.” – Cursor Research R&D Team
Beyond pre‑training, MoK is also suited for post‑training tasks, especially in reinforcement learning environments where determinism is crucial for reproducing experiences. Finance and healthcare industries, which demand strict model traceability, can benefit from this feature.
Conclusion and Call to Action
Mixture‑of‑Kittens represents a decisive step toward efficient MoE models. By making the megakernel open source, Cursor Research paves the way for broader adoption while maintaining cutting‑edge performance.
If you are a developer or researcher working with MoE models and your infrastructure meets the hardware requirements, download the Mixture‑of‑Kittens repository from GitHub today and start transforming your training pipelines. Feel free to share your feedback in the community to accelerate the evolution of this technology.
