Cursor unveils 'Mixture-of-Kittens' — 41% faster AI model training

Cursor, the essential coding editor for developers, has open-sourced a performance-optimizing infrastructure technology that is set to surprise everyone. It is called 'Mixture-of-Kittens,' a mega-kernel for training Mixture-of-Experts (MoE) models, specifically designed for NVIDIA's next-generation Blackwell NVL72 servers. They developed this core technology to push hardware to its limits to train their flagship AI coding agent, 'Composer,' and have decided to share it transparently with developers worldwide.

The 'Communication Bottleneck' no GPU can avoid

MoE models are highly performant because they are structured as multiple specialized models collaborating. However, they have a critical drawback: the 'all-to-all' communication process, where different GPUs must constantly exchange computation results, consumes more than half of the total training time.

This is similar to colleagues working together on a project. Even if each person finishes their task in a blink of an eye, what if they had to stop and exchange notes via a slow internal messenger to confirm every step before moving forward? Ultimately, they would waste more time sending notes and waiting for replies than actually doing the work.

Even with NVIDIA's latest Blackwell NVL72 system revolutionizing hardware transfer speeds, this barrier remained. Every time a GPU sent computation results, it had to undergo a synchronization step to match signals with the relatively slower main CPU. These tiny, recurring delays acted as a bottleneck, hindering the overall system.

A 'Mega-Kernel' that fuses communication and computation

To solve this, Cursor introduced the concept of a 'mega-kernel' that merges computation and communication into one. Instead of separating the computation stage that processes data from the communication stage that sends it to other GPUs, it is designed to process both simultaneously within a single pipeline.

This development actively leveraged PyTorch's symmetric memory technology. As a result, GPUs can access each other's memory directly via high-speed networks without having to go through the CPU, which is a relatively slower main processor.

Previously, there was a bottleneck where a GPU had to wait for the CPU to signal, "You may now send data," after completing its calculations. In the mega-kernel structure, this latency is completely eliminated. GPUs can read and write to each other's memory in real-time, allowing them to continue computations without interruption.

To run this mega-kernel, the following hardware and software requirements must be met.

# Mixture-of-Kittens 시스템 요구 사양 (2026년 8월 기준)
python --version      # Python 3.12 이상 필요
pip show torch        # PyTorch 2.10 이상 필요
nvcc --version        # CUDA 13.0 이상 및 Blackwell GPU 필요

By combining the potential of the latest software stacks with next-generation hardware to the extreme, they were able to achieve an optimization that essentially removes the traffic lights on the path where data flows.

'41% performance boost' proven in actual training

What were the actual results? The performance measurement data exceeded expectations.

In individual layer-level tests, MXFP8 forward pass, a next-generation data format, was up to 2.37 times faster than the fastest-known benchmark. It also recorded a 1.78 times speed increase for backward passes.

This performance gain wasn't limited to the lab. Cursor trained its top-tier coding model, Composer, on a cluster of 512 actual GB300 GPUs. As a result, the total training throughput of the final model increased by 41%.

A 41% increase in total training throughput in a large-scale server environment is a massive result. It proves that significant time and hardware costs for training AI models can be saved.

The 'Kittens' spread via open source

This exciting project didn't emerge from thin air. It was built upon 'ThunderKittens,' an open-source hardware acceleration framework developed by the Hazy Research team at Stanford University, which the Cursor research team expanded and completed. Behind the cute name 'Kittens' lies the persistent dedication of engineers aiming to squeeze every bit of performance out of GPU semiconductors.

This release resonates significantly within the AI industry. The competitiveness of AI service companies no longer stops at just using APIs made by others. 'Hardware-software co-design'—the ability to extract the absolute maximum potential from hardware architecture and build optimized software accordingly—is becoming the true differentiator.

Cursor has generously released this powerful mega-kernel technology under the Apache 2.0 license. This is lowering the technological barriers for leading companies handling large-scale, high-performance infrastructure and allowing more development teams to benefit from efficient MoE model training.

The true stage for performance competition lies in invisible infrastructure

Cursor's 'Mixture-of-Kittens' release is a prime example of how to maximize the potential of high-performance hardware. Of course, it will be interesting to see how the role of such complex mega-kernels changes when next-generation GPUs like the NVIDIA Rubin, which fine-tunes computation at the hardware level, arrive. But for now, this open-source tool will serve as a practical and reliable breakthrough for countless developers and research teams trying to overcome hardware limitations and boost model training speeds.