Perplexity Hybrid Compute is now live on Mac. The feature splits a single AI task between the cloud and the local model, so all your sensitive files stay on the machine.
“In the newest release of the Perplexity for Mac app, users can now add a compact local model post-trained for Perplexity Computer and choose to run tasks on device on their Mac,” noted Perflexity.
Cloud models handle the research and heavy reasoning, while a compact, local AI model on your Mac handles any step that touches private data.
Perplexity Hybrid Compute offers the best of both worlds
Running AI models locally offers advantages beyond privacy. Because the model runs directly on your Mac, there’s no need to send data to a remote server or wait for a round trip to the cloud, making certain tasks faster and more responsive. Local AI can also reduce reliance on (expensive) subscription-based cloud services.
The main downside is that local AI models generally aren’t as powerful as the biggest cloud-based models. They also depend on your Mac’s processor and memory, so demanding models can consume significant system resources and may run more slowly.
The goal of Perplexity Hybrid Compute is to offer the best of both these options.
All tasks start in the cloud, where a large model plans the work and does web research. But the moment it detects anything sensitive, the workflow shifts. Perplexity says it uses a classifier that flags names, addresses, account numbers and other sensitive information.
At this point, the privacy gate has four options. It can either process the task locally, mask the sensitive details, refuse the action or ask the user for consent. Perplexity claims that credentials, payment card numbers and government IDs get the strictest treatment of all.
This classifier is the real engine here. Perplexity says it built and trained the tool so it can catch personally identifiable information before it is sent to a server. Developed at its in-house Secure Intelligence Institute, the company also open-sourced the classifier.
Perplexity Hybrid Compute currently has three local models available. They include Google’s Gemma 4 E4B, Alibaba’s Qwen3.6 35B-A3B and a version of Qwen3.6 35B that Perplexity itself trained.
Easy to use
The setup takes one click, meaning you won’t have to touch Terminal or install a separate runtime.
Hybrid Compute also works across devices, meaning you can start a task from an iPhone or iPad. If your Mac is on and running the Perplexity app, it handles sensitive steps locally in the background.
Users looking for local inference on standby can also use a dedicated Mac mini to do the job. It stays on and can be remote-controlled using an iPhone.
Who Hybrid Compute is actually built for
Perplexity is aiming it at professionals who handle confidential information every day. In finance, a cloud model can assemble public filings and market comps for due diligence. The local model cross-references confidential deal documents and flags discrepancies with the web research.
At an ad agency, cloud research on audience trends can be checked against embargoed creative stored locally. The local model flags where the creative drifts from the data.
Lawyers can use cloud search to cover public case law. The local model extracts facts from privileged files and turns them into anonymized research questions for a brief.
Hybrid Compute system requirements
Perplexity Hybrid Compute requires an Apple Silicon Mac running on macOS 15 or later. Perplexity says it needs 24GB of unified memory as the baseline, but some reports point to 32GB as the recommended amount for larger local models.
The feature is currently rolling out for Perplexity’s Pro, Max and Enterprise subscribers through the Mac app. It comes days after Apple highlighted Perplexity’s Personal Computer as a use case for its new M6 Mac mini.
