Cursor Open-Sources Mixture-of-Kittens (MoK): A Deterministic MoE Training Megakernel for GB300 NVL72 Racks

Cursor Research has open-sourced Mixture-of-Kittens (MoK), the mixture-of-experts (MoE) training megakernel that powers its Composer models. MoK fuses all MoE communication and computation steps into a single deterministic kernel, achieving up to 2.37x higher throughput than the strongest public baseline. The kernel is already used to train Composer models across tens of thousands of GPUs, marking a significant step in scalable AI infrastructure. ## Deployability and Hardware Requirements MoK is available on GitHub under the Apache-2.0 license, but its hardware requirements are substantial. It requires NVIDIA Blackwell SM100 or SM103 GPUs—which means GB200 NVL72 or GB300 NVL72 rack systems—along with Python 3.12+, PyTorch 2.10+, and CUDA toolkit 13.0+. Inter-GPU communication relies on PyTorch's symmetric memory, which further limits deployment to environments with advanced memory management. Given these prerequisites, realistic adopters include frontier labs, well-funded model startups, GPU neoclouds, and national computing centers. Single-node teams and 8-GPU setups are not viable targets, as the kernel is designed for scale and deterministic performance. ## Technical Overview MoK's key innovation is its fully fused, deterministic design. By combining token dispatch, expert computation, and gradient reduction into one kernel, it eliminates synchronization overhead and reduces memory latency. This is especially critical in MoE models, where uneven expert loads often cause pipeline bubbles. The kernel's determinism ensures reproducible training runs, a valuable property for debugging and large-scale experiments. In 2026, as MoE models dominate frontier AI and hardware like the GB300 NVL72 becomes more accessible, tools like MoK set a new standard for efficient training. Its open-source release under Apache-2.0 invites broader innovation, though the high entry barrier means early adoption will remain concentrated among organizations with substantial compute resources. ## Implications and Outlook The release of MoK is a notable milestone in the ongoing push for efficient AI infrastructure. As Cursor continues to deploy this megakernel in production, the community gains a reference implementation that could influence future training frameworks. However, the hardware floor ensures that its immediate impact is felt primarily in high-end AI research and commercial environments, while smaller teams will likely wait for more accessible versions or further optimizations.

via MarkTechPost