Perplexity open-sources Lily: 1.35x faster than MLX-LM on Apple Silicon

Perplexity has open-sourced Lily, the local inference engine powering hybrid compute in its Perplexity Computer Mac app. Lily is built in Rust with hand-written Metal kernels and contains no PyTorch or MLX in the execution path. It runs a single model, Qwen3.6-35B-A3B, on Apple Silicon hardware, with Perplexity stating that narrowness is the performance argument. Benchmarks on an Apple M5 Max with 128 GB RAM showed Lily averaging 4,156 prefill tokens/s versus MLX-LM's 3,388 (1.23x faster), and 170.0 decode tokens/s versus 126.4 (1.35x faster). A teacher-forced correctness check across 192 token positions showed Lily's perplexity only 0.04% higher than the reference, with identical top-ranked token 96.35% of the time, indicating quality parity. Speculative decoding was tested but found to slow inference by 18%. Lily exposes an OpenAI-compatible chat-completions API and the source code is published under the perplexityai GitHub organization.