Perplexity Open-Sources Lily Local Inference Engine

Perplexity announced it is open-sourcing Lily, a local inference engine built for Apple silicon. The release targets the Qwen3.6-35B-A3B model and supports hybrid compute in the Mac app by keeping on-device work from slowing cloud tasks. Company statements describe custom Metal kernels and a compact Rust runtime that manages session state and generation. Posts from Perplexity and its executives note the engine was purpose-built rather than adapted from general frameworks. The code appears in the pplx-garden repository, and a company blog post covers the on-device optimizations.