Analysis
Perplexity introduced hybrid compute for its Mac app on Sept. 1, a mode that splits each Perplexity Computer task between frontier models running in the cloud and a local model running on the Mac itself, Unite.AI reported. Under the new architecture, the cloud handles frontier reasoning, web search and planning, while the local model processes private files, sensitive information and on-device actions -- keeping anything sensitive from leaving the machine by default rather than relying on a post-hoc redaction step.
The next day, Sept. 2, Perplexity open-sourced Lily, a local inference engine written in Rust with hand-written Metal kernels built specifically for Apple Silicon, MarkTechPost reported. On the Qwen3.6-35B-A3B model, Perplexity's own benchmarks show Lily running 1.23x faster on prefill throughput and 1.35x faster on decode throughput than Apple's own MLX-LM framework -- a direct challenge to Apple's on-device inference tooling, published by a company that doesn't make the hardware it's optimizing for.
Hybrid compute ships with three components:
“- Three local models at launch -- Gemma 4 E4B, Qwen3.6 35B-A3B, and a Perplexity model post-trained specifically for Perplexity Computer.”
- Lily -- the local inference engine itself, open-sourced and benchmarked against MLX-LM.
- Privacy Gate -- a small on-device classifier, open-sourced on Hugging Face, that inspects outbound data for PII such as names, addresses and account numbers and lets the user decide before anything sensitive is sent to the cloud.
- Three local models at launch -- Gemma 4 E4B, Qwen3.6 35B-A3B, and a Perplexity model post-trained specifically for Perplexity Computer.
The releases land the same week Nvidia's $12.93 billion acquisition of Hugging Face raised questions about whether the platform stays neutral once a chipmaker owns it. Perplexity choosing to publish safety-relevant tooling like Privacy Gate on Hugging Face anyway, the same week that deal closed, is a practical vote of confidence that the hub remains usable regardless of ownership -- a real-world signal that matters more than any acquirer's public pledge about staying compute-agnostic.
For AI application developers, Lily's benchmark claims against MLX-LM are worth independent verification before building on them -- vendor-published throughput comparisons on a vendor's own benchmark are common this cycle, and Apple hasn't yet responded publicly. The more durable signal is architectural: on-device PII screening before cloud calls is becoming a genuine product category, not just a compliance checkbox, and any startup building AI products that touch sensitive user data should be evaluating whether a similar local-gate step belongs in its own pipeline.
