Perplexity open sourced Lily, a Rust and Metal inference engine running Qwen3.6-35B-A3B on Apple silicon, 1.35x faster decode.