analiza aiqbench

Perplexity's open-sourced Lily engine demonstrates that specialized Rust and Metal kernels can significantly outperform general-purpose MLX stacks on Apple Silicon for specific MoE models. By optimizing exclusively for Qwen3.6-35B-A3B, this runtime achieves higher throughput by bypassing generic frameworks like PyTorch or standard MLX in favor of hand-written operations tailored to the model's unique activation patterns. This approach highlights a strategic trade-off where sacrificing cross-model versatility yields measurable performance gains for local inference on high-end Mac hardware.