Perplexity Open-Sources 'Lily' Inference Engine for Apple Silicon
Perplexity AI has open-sourced "Lily," a specialized Mac inference server engineered to maximize execution performance for Qwen 3.6 models on Apple Silicon hardware. The project was published in Perplexity's official GitHub repository to provide dedicated, highly optimized on-device inference. By tailoring optimizations to a specific model architecture and Apple Silicon hardware, Lily enables Mac users to run local inference at significantly higher speeds without relying on general-purpose frameworks. This release advances the adoption of hybrid local-cloud workflows, providing lower latency and enhanced privacy for on-device AI operations. Lily features separate optimizations for prompt prefilling and token generation tailored specifically for the Qwen 3.6 model family (such as Qwen3.6-35B-A3B). Rather than acting as a multi-model server like Ollama, it focuses exclusively on single-model efficiency to extract maximum performance from Apple Silicon unified memory.
## BACKGROUND
Local LLM inference on consumer hardware relies heavily on unified memory architectures like Apple Silicon, which grant GPUs direct access to high-bandwidth system RAM. Qwen is an open-weights large language model family developed by Alibaba Cloud that is widely used across the AI community. While general-purpose engines like Ollama support many architectures, hyper-specialized inference servers can unlock superior compute and memory efficiency for specific hardware-model pairings.