You want AI to work on your files, but you don’t love the idea of feeding your private stuff into a cloud model you don’t control. That tension is exactly what Perplexity Hybrid Compute is trying to solve.
It’s a new feature in the Perplexity Mac app, and it does something most AI tools won’t. It splits every task in two: the cloud handles the heavy reasoning, while a local model on your own Mac handles anything sensitive. Here’s how Perplexity Hybrid Compute works and whether it’s worth setting up.
What is Perplexity Hybrid Compute?
Perplexity already has a product called Perplexity Computer, a suite of AI agents that can autonomously complete tasks using the web, your files, and your apps. Perplexity Hybrid Compute is the new twist on that.
Instead of sending your whole task to the cloud, it splits the work between a big cloud model and a small local model running right on your Mac. Cloud models handle research, reasoning, and web search. The local model deals with tasks that touch private files and sensitive information.
The whole point, in one sentence: your sensitive data stays on your device.
How a task gets split
Say you ask it to prepare a document that combines your research with a confidential client file. Here’s what actually happens behind the scenes.
The cloud model starts the task. It does the web research, the planning, the reasoning. Then, when it needs to touch your private files, it hands that part off to the local model on your Mac. No sensitive content crosses to the cloud.
Perplexity gives a concrete example: a lawyer preparing a brief that compares a current case against existing case law. The public research can use the cloud. The client’s confidential information stays local.
There’s also a cost angle. Local model work doesn’t use cloud credits. You’re not charged for tokens a local model generates on your own machine. For heavy users, that can add up to real savings.
The privacy gate and PII classifier
This is the clever part of Perplexity Hybrid Compute, and it’s worth understanding.
Before anything leaves your Mac, an on-device PII classifier reads the task. It looks for names, addresses, account numbers, and other personal information. When it finds something sensitive, it swaps those details for stand-ins before the request is sent to the cloud, then restores them when the answer returns.
It’s not a black box either. A “privacy gate” on your Mac controls what’s allowed to leave. You can keep a file local, mask the sensitive bits, refuse the action entirely, or give consent to send it anyway.
Perplexity also open-sourced the classifier, which it says it trained with its Secure Intelligence Institute. Open-sourcing it is a nice credibility move, since it means independent researchers can actually inspect how the privacy detection works.
What you need to run it
Here’s the catch, and it’s a real one. This isn’t for every machine.
You need an Apple Silicon Mac running macOS 15 or newer. Perplexity says 24GB of unified memory is the minimum, with 32GB recommended. If you’re on an 8GB or 16GB Mac, it won’t run.
The feature is available to Pro, Max, and enterprise customers. So it’s not a free feature for everyone.
On the plus side, the local models install with one click. You don’t open a terminal, you don’t mess with Ollama or any manual runtime setup. Options include Gemma E4B and a couple of flavors of Qwen’s 3.6 model, including one post-trained by Perplexity.
How to set it up
If your Mac qualifies, the setup is refreshingly simple.
Step 1: Update Perplexity for Mac. Download and open the latest version of the app.
Step 2: Download a local model. Pick one from the in-app list and install it in one click. No terminal, no API key, no manual runtime.
Step 3: Choose Hybrid. Open the model selector and pick Hybrid, then choose your local and cloud models for the task.
Step 4: Run a task. As you work, Perplexity will scan what it’s asked to handle. If it detects private information, it’ll ask what you want to do: keep it local, mask it, or let it through.
That last step is the part that might surprise you. You get asked, you don’t just silently trust it.
Who it’s for
Use it if you work with sensitive data and want AI to help without shipping everything to the cloud. Think freelancers with client NDA material, lawyers, HR people, anyone handling financial or personal records.
It’s also for thrifty users. Offloading work to a local model saves cloud credits, and on a subscription where you’re watching usage, that matters.
Skip it if you’re on a lower-RAM Mac, or if you’re happy sending everything to the cloud and don’t handle particularly sensitive files. If you don’t need the privacy gate, it’s complexity you don’t have to add.
The bottom line
Perplexity Hybrid Compute is a thoughtful answer to a real problem. Keeping AI powerful while keeping your private data on your own device is genuinely hard, and this is one of the cleaner attempts at it.
The hardware requirement holds it back from being for everyone. But if you have a capable Mac and you’ve been uneasy about feeding sensitive files to a cloud model, Perplexity Hybrid Compute is worth a serious look. It’s also a nice preview of where AI tools are heading, not everything has to leave your machine.
If you’re curious about running AI locally more broadly, we’ve covered how to run an LLM on a tiny microcontroller and ways to cut your AI costs. Privacy is becoming a real selling point, and this is one of the better examples of it. You can also check Perplexity’s own walkthrough to see the setup for yourself.
