Summary & Key Takeaways ​

• Cursor has open-sourced a new MoE megakernel named "Mixture-of-Kittens."
• It is specifically designed for Mixture-of-Experts (MoE) models.
• The megakernel targets NVL72s hardware.
• This release provides specialized infrastructure for high-performance AI.

Our Commentary ​

"Mixture-of-Kittens" is a fantastic name, first off. An open-source MoE megakernel for NVL72s sounds like a deep dive into the hardware-software co-design for large AI models. This is the kind of foundational work that enables future breakthroughs. It's not for everyone, but for those pushing the limits of AI performance, this is a big deal.