Perplexity
@perplexity_ai
Perplexity is **open-sourcing Lily**, the local inference engine behind hybrid compute in Perplexity Computer.
- Specialized for **Qwen3.6-35B-A3B** on **Apple silicon** so on-device steps don’t bottleneck Computer tasks.
- Custom Rust + Metal stack (no PyTorch/MLX in the execution path) with separate prefill and decode optimizations.
- On an M5 Max MacBook Pro, Lily averaged **1.23×** MLX-LM prefill and **1.35×** decode throughput across ten lengths (256–128K), with output quality effectively unchanged.
- Code available on GitHub (`perplexityai/pplx-garden` / Lily); engineering blog details the optimizations.
Official Perplexity engineering blog + @perplexity_ai announcement thread.
