Perplexity Open Sources Lily: Rust + Metal Inference Engine for Qwen3.6-35B on Apple Silicon

Perplexity has open-sourced Lily, a high-performance inference engine written in Rust and using Apple's Metal GPU API, specifically optimized to run Qwen3.6-35B-A3B on Apple Silicon hardware. The engine is designed to maximize throughput on M-series chips without relying on PyTorch or standard ML runtimes, making it a low-level, production-grade alternative for local inference on Macs. For developers building on Apple Silicon — whether for privacy, latency, or cost reasons — Lily offers a path to running a 35B-class MoE model locally with significantly better hardware utilization than general-purpose frameworks. The Rust implementation also makes Lily a strong foundation for developers who want to contribute to or fork a systems-level inference stack. This release adds meaningful competition to the growing ecosystem of Apple Silicon inference engines alongside tools like llama.cpp and MLX.