According to a recent LinkedIn post from Cursor, the company is open-sourcing its Mixture-of-Kittens (MoK) training megakernel for NVL72 GPUs. The post indicates that MoK combines all Mixture-of-Experts communication and computation in a single deterministic kernel and reports performance up to 2.37x faster than leading public baselines.
The LinkedIn post further notes that MoK is already used to power training across tens of thousands of GPUs at Cursor, where it has increased end-to-end training throughput by 1.41x compared with a prior DeepEP-based stack. The content suggests Cursor is investing in infrastructure efficiency while positioning itself as a contributor to the wider AI research ecosystem.
For investors, the post implies improved internal compute efficiency, which may support faster model iteration and potentially lower unit training costs over time. By open-sourcing MoK, Cursor could strengthen its reputation among AI practitioners, attract technical talent, and increase strategic optionality for collaborations, though direct near-term revenue impact is not clearly indicated.
The emphasis on scaling across large GPU fleets highlights Cursor’s focus on high-performance training environments, which may be relevant if the company competes in frontier or enterprise AI models. If MoK gains adoption beyond Cursor, the company could benefit indirectly through ecosystem influence and network effects, reinforcing its position in advanced AI infrastructure and tooling.