Perplexity launches Hybrid Compute for Mac to keep sensitive AI tasks on-device

The feature is part of Perplexity's broader Computer suite, which debuted in February as a set of AI agents that can complete tasks using the web, files, and apps. Perplexity first teased the hybrid approach in June, and the company has now brought it to Mac users.

According to Perplexity, each task starts in the cloud, but when triggered from an iPhone, the Mac accesses files and runs sensitive steps locally. The company says an on-device personal-information classifier reads each task before it is sent; names, addresses, and account numbers are swapped for placeholders and restored when the answer returns. Perplexity said it open-sourced that classifier, which was trained with its Secure Intelligence Institute.

Engadget reported that Hybrid Compute can split a task between cloud models such as Opus 5 or GPT-5.6 Sol and a local LLM, keeping sensitive data on the machine. Perplexity suggested a lawyer preparing a brief could keep client data confidential by running that part locally, and said offloading some work from more expensive frontier models could save users money.

Jon Staff, who oversees Perplexity's Mac products, told Engadget that the feature is integrated directly into the Mac app. Whenever a user tries to upload files or send information, the app automatically checks for sensitive content and confirms before sharing data to the cloud. A new privacy classifier suggests files and information that should stay on the computer, and users can review those suggested files before the system delegates tasks between models.

Users can also choose which local models to use. Engadget said the options are Gemma E4B and two variants of Qwen's 35-billion-parameter 3.6 model, one of which was post-trained by Perplexity. 9to5Mac, meanwhile, described a downloadable model called PPLX Qwen 3.8 27B. Both outlets said installation is one-click and does not require opening the Mac's terminal or using Ollama. Perplexity said more local models will be added in the future.

During execution, the Mac app displays local CPU, GPU, and memory usage, with a sidebar showing token consumption. Tokens generated locally are not charged against a user's cloud credits. After the system produces an output, users can give follow-up instructions, and tasks can also be queued from an iPhone.

Hybrid Compute requires Apple Silicon with macOS 15 or later. Perplexity's website lists 24GB of unified memory as the minimum and 32GB for best results, but Engadget reported the company recommends at least 32GB. The feature is available to Pro and Max subscribers as well as enterprise customers.

Staff acknowledged that a fully cloud-based frontier model will almost always produce better raw output, saying such models are more expensive and more capable. But he said some users do not need the most powerful AI for their work, and for them privacy and cost may matter more. He described the tradeoff as a sliding scale, with the goal of giving users control over which models handle a given task.

Perplexity has been expanding its Mac presence this year. In May it overhauled its Mac app, and Apple has cited Perplexity Personal Computer as a productivity use case for the new M6 Mac mini.