PPLX CUDA KERNELS

Summary from the TLDR AI newsletter

Perplexity has released some of its MoE kernels, which outperform DeepSeek at scale while being slightly more flexible and less opinionated about the MoE architecture.