Perplexity Hybrid Compute: the PII never leaves the Mac

Perplexity shipped Hybrid Compute for its Mac app today. The headline is cost — work done locally burns no cloud credits — but the interesting part is the boundary. An on-device classifier reads each task before it leaves the machine and decides what the cloud is allowed to see. That is a design pattern worth stealing whether or not you ever install this app.

What actually happened

A task starts in the cloud. Before it goes, a local PII classifier scans it and substitutes stand-ins for names, addresses and account numbers; steps that touch genuinely sensitive material run against a model on your Mac instead. You get a prompt to review what was flagged before anything executes (9to5Mac, MacStories).

The hardware floor is Apple silicon, macOS 15 or later, 24GB of unified memory minimum with 32GB recommended. Local model options at launch include a one-click PPLX Qwen 3.8 27B, plus Gemma E4B, Qwen3.6 35B-A3B and a Perplexity post-trained Qwen3.6 35B. No Ollama install, no runtime setup, no API key. It is gated to Pro, Max and enterprise plans. Perplexity says the classifier itself was open-sourced through its Secure Intelligence Institute work.

This is a different product from the Portable Computer release last week, which ran everything locally on Nvidia hardware. Hybrid Compute keeps the frontier model in the loop and moves only the sensitive fraction down to the laptop.

Why a local PII boundary matters for your business

Most small operators have already made this decision by accident. Somebody pasted a customer list into a chat window to get it reformatted. Somebody dropped a vendor invoice into an assistant to extract line items. Nobody logged it, nobody approved it, and the data is now in a retention policy you have never read.

A classifier that runs before the network call is the right shape for the fix, because it does not depend on anyone remembering a rule. The 27B-class local models are perfectly competent at the work that touches real customer records — extraction, redaction, classification, matching — and the frontier model still gets the reasoning-heavy leftovers. Route by sensitivity, not by vibe.

The part to be sober about: this boundary lives inside one vendor's app, on Macs with 24GB of RAM, behind a subscription. That is fine for a founder's laptop and useless as a company policy. If the pattern matters to you, the durable version is the same two-stage pipeline in your own stack — a small local model doing redaction, a config value naming whichever cloud model handles the rest. Same result, and it survives the day Perplexity changes its pricing page.

Sources: 9to5Mac, MacStories.