Perplexity is putting part of its AI on the Mac. On September 1, 2026, the company introduced Hybrid Compute for its Perplexity Mac app, splitting Perplexity Computer tasks between frontier models in the cloud and a local model running on the user’s Mac. The result is a choice between local processing and cloud capability without forcing every request onto one side of the fence.
Hybrid Compute is available to Perplexity Pro, Max, and Enterprise subscribers. It runs on Apple silicon Macs with macOS 15 or later and at least 24GB of unified memory, although Perplexity recommends machines with at least 32GB. The company called Apple’s new Mac mini and Mac Studio ideal platforms for the setup.
The launch includes three local models: Gemma 4 E4B, Qwen3.6 35B-A3B, and a Perplexity model post-trained for Perplexity Computer. Users install the latest Perplexity Mac app, download a local model, open the model selector, choose Hybrid, and select the local and cloud models they want to use. No terminal ritual is required; the app handles installation.
Local processing with a privacy gate
Perplexity’s main argument for Hybrid Compute is control. Work handled by the local model uses no cloud credits, and tokens generated on the user’s machine are not charged. Once a task starts, the app shows CPU, GPU, and memory usage alongside token consumption, giving users a view of what their Mac is actually doing.
The system also includes a privacy gate on the Mac that controls what can leave the device. It can mask sensitive details, keep content local, refuse actions, or ask for user consent before sending information to the cloud.
“This is integrated directly into the Mac app. Any time you try to upload files or send information, we’re going to automatically check for sensitive content and make sure that you want to share that data to the cloud,” said Perplexity’s Jon Staff, who oversees all of the company’s Mac products.
The gate uses an on-device Perplexity classifier called PII-Tracer to identify names, addresses, account numbers, and secrets. Credentials, payment card numbers, and government IDs receive the strictest protection: detected values are replaced with stand-ins before a request leaves the Mac, then restored when the answer returns.
Perplexity described PII-Tracer as a 0.6B-parameter bidirectional encoder adapted from a Qwen3 backbone. The company also introduced PII-TRACE, a benchmark built from 13,148 synthetic user-assistant conversations across 13 languages. Privacy tooling is not the most glamorous part of an AI launch, but it is the part that decides whether sensitive files stay sensitive.
A trade-off between capability and cost
Hybrid Compute follows Perplexity Computer, which debuted in February as a suite of AI agents that can autonomously complete tasks using the web, files, and apps. Perplexity’s platform now includes Perplexity Computer, Personal Computer for the Mac, and Hybrid Compute—a progression from cloud agents toward a setup that can keep selected work on the device.
The company presents Hybrid Compute as useful for keeping client data confidential and saving inference costs. Users can also run Computer on a dedicated, always-on Mac mini controlled from an iPhone, turning a small desktop into a persistent local gateway for computer-based tasks.
There is a clear compromise. Hybrid Compute’s output is less capable than fully cloud-based models, which can matter when a task demands stronger reasoning or more polished artifact creation. “The short answer is that a fully frontier output is going to almost always be better in terms of raw artifact creation. It’s more expensive, it’s more capable,” Staff said.
That trade-off is the point rather than a flaw Perplexity is hiding behind a product label. “I think it’s got to be a sliding scale, and we want the user to have control over where they are on that sliding scale for the particular type of work they want to do,” Staff said. Hybrid Compute gives Mac users that scale, provided they have compatible hardware and can accept that local privacy comes with a ceiling on capability.
Based on
