Perplexity AI—Perplexity AI open-sourced pplx-kernels, a high-performance MoE communication library, under MIT license

Perplexity released pplx-kernels, a library of GPU communication kernels for mixture-of-experts (MoE) model inference, as open-source software under the permissive MIT license, publishing the code on GitHub with an accompanying engineering blog post. Perplexity states the library achieves substantially faster all-to-all communication than standard primitives and supports multiple hardware transports (NVLink, IBGDA, IBRC, AWS EFA). The repository has drawn outside contributors and underpins a published MLSys paper. This is part of a broader, ongoing pattern of Perplexity open-sourcing ML infrastructure under MIT licenses (also including pplx-garden and the pplx-embed embedding models with public technical reports), which is a genuine contribution to the open-source ML ecosystem distinct from the company's product/business controversies.