Loading Perplexity Benchmarks Its Open-Source Engine Lily: 1.23× Prefill and 1.35× Decode Gains vs MLX-LM - AI Pulse...