Perplexity Launches Hybrid Compute for Mac, Splitting AI Tasks Between Cloud and Local Models

Perplexity has launched Hybrid Compute, a new feature in its Mac app that splits AI tasks between cloud models and a local model running on-device. The rollout follows an initial announcement of the work in June and arrives after Apple featured Perplexity's Personal Computer app as a productivity use case for the new M6 Mac mini.

How it works

According to Perplexity, its "Computer" assistant now starts each task in the cloud but can shift to a local model for steps that touch private files or sensitive data. A task triggered from an iPhone, for example, can hand off to the Mac to access local files and complete sensitive processing without that data leaving the device.

Privacy protection is handled by an on-device PII classifier that scans each task before anything is sent to the cloud. Names, addresses, and account numbers are replaced with placeholder values, then restored once the cloud model returns its answer. Perplexity says it developed the classifier with its Secure Intelligence Institute and has open-sourced it, though the company has not published independent benchmark data verifying its accuracy at detecting or masking personal information.

Hardware requirements

Hybrid Compute only runs on Apple silicon Macs with macOS 15 or later. Perplexity sets 24GB of unified memory as the minimum requirement and recommends 32GB for best results — meaning Macs with 8GB or 16GB of RAM cannot run the feature at all. This effectively limits Hybrid Compute to higher-end MacBook Pro, Mac Studio, and upper-tier Mac mini configurations.

Setup requires a one-click download of what Perplexity calls "PPLX Qwen 3.8 27B," a local model built on Alibaba's Qwen 3.8 architecture at 27 billion parameters. Perplexity says the process requires no manual runtime configuration or third-party tools like Ollama. Tasks handled entirely by the local model consume no cloud credits and require no API key, according to the company.

What this means

Hybrid Compute is a product feature, not a new foundation model — the underlying local component is a repackaged version of an existing Qwen checkpoint, not a Perplexity-trained model from scratch. Its real significance is architectural: Perplexity is betting that a split-execution model, where sensitive data never leaves the device while general reasoning still runs in the cloud, becomes a meaningful differentiator as AI assistants gain deeper access to personal files.

The 24GB memory floor is a hard constraint that excludes a large share of the existing Mac installed base, including every base-model MacBook Air and many entry Mac mini configurations sold in the past two years. That narrows Hybrid Compute's near-term audience to higher-end hardware buyers — likely by design, since Apple's own promotion of Perplexity alongside the M6 Mac mini suggests the two companies are coordinating on positioning higher-memory Macs as AI-capable machines. Whether the open-sourced PII classifier holds up to independent scrutiny will determine if this privacy pitch is substantive or largely marketing.