# Optimizing On-Device Inference for Apple Silicon

By Steven Van · 2026-09-01

Perplexity describes Lily, a lightweight local inference engine built for Apple silicon and Qwen3.6-35B-A3B, with separate optimizations for prefill…

[Perplexity](<https://creatorstoolbox.com/tools/perplexity>) has built Lily, a lightweight local inference engine for running Qwen3.6-35B-A3B on Apple silicon, with separate optimizations for the prefill and decode stages.

The [announcement](<https://www.perplexity.ai/hub/blog/optimizing-on-device-inference-for-apple-silicon>) walks through how Qwen's architecture opens up model-specific optimizations on Apple silicon, where further tuning stops paying off, and gives an end-to-end comparison against MLX-LM. Lily is to be open-sourced.

![Perplexity](<https://www.google.com/s2/favicons?domain=perplexity.ai&sz=128>)

Perplexity

AI-powered answer engine — ask anything and get cited, real-time answers from across the web.

[View Perplexity →](<https://creatorstoolbox.com/tools/perplexity>)

Sources

[Original source](<https://www.perplexity.ai/hub/blog/optimizing-on-device-inference-for-apple-silicon>)

[Read Optimizing On-Device Inference for Apple Silicon on Creators Toolbox](<https://creatorstoolbox.com/blog/perplexity-optimizing-on-device-inference-for-apple-silicon>)

---
Canonical source: https://creatorstoolbox.com/blog/perplexity-optimizing-on-device-inference-for-apple-silicon