Perplexity just dropped the source for Lily, the inference engine underneath Hybrid Compute in Perplexity Computer, and the numbers are worth paying attention to. On a 40-core, 128 GB M5 Max, Lily hit 4,156 prefill tokens per second against MLX-LM’s 3,388 — that’s a 1.23x lead. Decode was even better: 170 tokens per second versus 126.4, or 1.35x faster. Every single test point from 256 to 128K tokens came out ahead.

What makes Lily interesting isn’t just the speed. It’s the deliberate refusal to be general. There’s no PyTorch or MLX anywhere in the execution path. A Rust layer loads the checkpoint and runs the generation loop, an OpenAI-compatible chat-completions API streams tokens, and hand-written Metal kernels do the heavy lifting. The engine supports exactly one model — Qwen3.6-35B-A3B — on exactly one hardware family, Apple silicon. That narrowness is the whole argument.

The architecture squeezes performance from three places. Prefill gets a 77.4% boost from fusing 4-bit dequantization directly into the grouped GEMM, so expanded bfloat16 weights never touch unified memory. Keeping expert routing, prefix scan, scatter, and block map inside a single GPU command buffer added 89% at 512 tokens. Decode wins come from coalesced cache reads — key bandwidth jumped from 33.8 to 47.9 GB/s — plus GQA packing that improved decode 23.8% at 32K context, and a fixed-block attention layout that delivered 40.2% at 128K.

Perplexity isn’t asking anyone to take the numbers on faith. The demo is public in the pplx-garden repo: a Rust and Metal inference server with greedy text generation behind a minimal HTTP API. The 4-bit checkpoint is 19.4 GB, so a Mac with 32 GB or more of unified memory is the realistic floor — Perplexity’s own shipping product lists macOS 15+, 24 GB minimum, 32 GB for best results. A teacher-forced check across 192 positions found Lily’s perplexity within 0.04% of the reference, with the same top-ranked token 96.35% of the time.

This is the same playbook vLLM and TensorRT-LLM ran on NVIDIA hardware: give up portability, assume one architecture, and turn every freed degree of freedom into speed. For Apple silicon users who’ve been waiting for a local inference stack that doesn’t feel like a compromise, Lily is the strongest signal yet that the Mac is a first-class citizen rather than an afterthought.