Perplexity launches hybrid compute on Mac: cloud and local models to protect sensitive data

Perplexity has unveiled "hybrid compute on Mac", a feature for its Perplexity Computer app on Mac that combines cloud processing with AI models running locally on the machine itself. The underlying idea is simple: the more powerful models in the cloud handle complex reasoning, web search and planning, while a compact model running on the Mac processes the files, tools and information the user would rather not send off the device. The addition extends "Personal Computer", the feature announced in April that already allowed Perplexity Computer to work locally with a Mac's files and tools; now it adds the ability to install a local model post-trained specifically for Perplexity Computer and to decide, task by task, whether it runs on the device or in the cloud.

The central mechanism is what the company calls the "privacy gate": before any information from a protected file reaches the cloud, this filter can mask the sensitive details, keep the data local, reject the action outright or ask for the user's explicit consent. As the article describes, Computer orchestrates the entire workflow automatically, so the user does not have to manually split their requests into fragments or move information from one tool to another: the split between cloud and device happens transparently within a single task.

The piece illustrates the approach with three examples from sectors where "important" work almost always involves something sensitive. In financial services, Computer can gather public information in the cloud (regulatory filings, comparables, market data) while the local model on the Mac cross-references that data with confidential deal documents and spots discrepancies between the two. At an advertising agency, research into cultural trends and audience demographics happens in the cloud, and the local model compares those findings with unreleased creative work and assets that reside only on the Mac. And in the legal field, a lawyer could research case law on the web while the local model summarizes privileged documents, turns those summaries into anonymized research questions and updates a draft brief with the findings, without attorney-client protected information ever leaving the machine.

To decide what should stay on the device, the privacy gate incorporates a Perplexity classifier that runs on the Mac itself and identifies sensitive data such as names, addresses, account numbers or secrets before they can leave the machine. Different levels of protection are applied based on that classification: the most critical data —credentials, payment card numbers or government IDs— receives the strictest treatment, and the privacy gate can choose to keep it local, reject the action entirely or rewrite the request so the cloud model can keep working without needing access to that protected information.

The feature also arrives with controls designed for businesses. Perplexity Enterprise subscribers will be able to define organization-level rules on what information must always stay on the Mac, what can be masked before being sent to the cloud and what actions require explicit user approval before leaving the device. Administrators will also be able to audit when and what information leaves a machine, giving legal, healthcare, financial and other regulated departments a consistent way to use Computer for tasks involving sensitive data.

At launch, hybrid compute includes three local models to choose from: Gemma 4 E4B, Qwen3.6 35B-A3B and a Perplexity in-house model. The feature requires a Mac with Apple silicon, macOS 15 or later and at least 24GB of unified memory, a threshold that, according to the article, allows a wide range of users to take advantage of local inference on hardware they already own. The article also notes that Apple's new Mac mini and Mac Studio, announced the previous week, are especially well suited platforms for this use thanks to the performance and unified memory of Apple silicon. Those who want access to local inference from anywhere can dedicate an always-on Mac mini and control it remotely from an iPhone, since Perplexity Computer is integrated with the phone: the user can send instructions from anywhere while the local model processes their most sensitive data privately and entirely on the device.

The feature is already available to subscribers of Perplexity's Pro, Max and Enterprise plans, and is enabled from the Mac app by downloading a local model in one click and selecting hybrid mode in the model picker, where the user chooses which local model and which cloud model to use for each task. Overall, the announcement reinforces Perplexity's push to position Computer not just as a cloud search and reasoning assistant, but as an orchestration layer that can handle professional work involving confidential information without forcing users to choose between AI power and data confidentiality.