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

Cursor Research has announced the open-source release of Mixture-of-Kittens (MoK), a specialized training kernel designed to significantly accelerate the development of mixture-of-experts (MoE) AI models. This development represents a substantial contribution to the machine learning community's efforts to optimize training efficiency on advanced hardware infrastructure. By open-sourcing this technology, Cursor is enabling broader adoption of advanced training techniques previously limited to well-resourced organizations.

Mixture-of-Kittens operates as a deterministic megakernel that consolidates all mixture-of-experts communication and computational operations into a single unified process. This architectural innovation eliminates the inefficiencies typically associated with managing multiple separate operations across distributed systems. When tested on NVIDIA's GB300 NVL72 racks—among the most advanced AI training hardware available—MoK demonstrated performance improvements of up to 2.37x compared to the strongest existing public baseline solutions.

The kernel was originally developed to power Cursor's Composer models and has now been made publicly available to accelerate innovation across the AI research and development community. This performance advantage stems from MoK's ability to reduce communication overhead and synchronization delays that normally plague large-scale distributed training operations.

• Democratizes access to optimized MoE training techniques previously available only to well-funded research labs
• Reduces training time and computational costs for organizations developing large mixture-of-experts models
• Enables researchers to experiment more rapidly with MoE architectures on cutting-edge hardware
• Provides a reference implementation that could influence industry standards for distributed AI training
• Potentially accelerates development of more efficient large language models and multimodal systems

The open-sourcing of Mixture-of-Kittens addresses a critical bottleneck in AI model development: the efficiency of training specialized architectures on expensive infrastructure. As mixture-of-experts models become increasingly central to scaling large language models and other advanced AI systems, optimizing training performance directly reduces both computational waste and development timelines. By sharing this technology publicly, Cursor contributes to a more efficient ecosystem where innovation benefits from shared infrastructure improvements rather than proprietary optimization silos.