According to a Hugging Face blog post, the team used Anthropic's Claude to build CUDA kernels, the low-level programs that run on NVIDIA GPUs. The kernels were then used to teach open models how to write similar code, effectively turning the proprietary model's output into a curriculum for open-weight models.
CUDA kernels are a specialized area of programming, and generating them with an LLM is a different task from writing ordinary code. The Hugging Face effort appears to be an attempt to close that gap for open models by letting them learn from an existing frontier model.
With a single source, there is no external comparison to make. The headline itself makes the only claim: Claude built CUDA kernels, and those kernels were used to teach open models.