Perplexity says its Lily AI engine runs Qwen faster than MLX-LM on Apple silicon. The specialized local inference engine will eventually become open source.