Perplexity's new hybrid AI keeps your sensitive files on your Mac while still using cloud intelligence
A built-in privacy filter decides in real time which parts of a task stay local and which go to the cloud. Lawyers, finance teams and healthcare workers are the target audience.
Key points
- Perplexity launched hybrid compute for its Computer agent platform, available to Pro, Max and enterprise subscribers on Apple silicon Macs running macOS 15 or later.
- An on-device classifier scans for personally identifiable information before any data leaves the machine, giving the user final say over what gets shared.
- Tokens processed locally cost no credits; users pay credits only on the cloud orchestration portion of a task.
- Three local models are available at launch: Google's Gemma E4B, Alibaba's Qwen3.6 35B-A3B, and a Perplexity-tuned version of Qwen3.6 35B.
- Perplexity says it isn't using this data for model training globally, though specifics for non-enterprise accounts haven't been published yet.
Most AI agents work like a postal service that opens your letters. You hand over everything, the cloud reads it, and the answer comes back. Perplexity wants to change that.
The company launched what it calls hybrid compute: a system where a single AI agent, software that can carry out multi-step tasks on its own, splits its workload between powerful cloud servers and a smaller model running entirely on the user's own Mac. Sensitive data never leaves the device.
How does the privacy filter actually work?
A piece of software Perplexity calls the Privacy Gate sits on the Mac and scans everything before it goes anywhere. It's a classifier, a trained program that reads text and decides what category it belongs to, and its job is to spot personally identifiable information: names, addresses, account numbers, anything confidential.
When it flags something sensitive, it stops and asks you. You decide whether that portion runs locally or gets sent out. If you choose local, those words never touch a cloud server.
Cloud handles the bigger-picture thinking: web research, long-range planning, heavy reasoning. Local handles whatever touches private files or actions on the machine itself.
"The cloud orchestration will break down the task based on the prompt and figure out how to route it to different subagents," Jon Staff, who leads Perplexity's macOS and iOS engineering, said at a press briefing first reported by VentureBeat. "That portion of the task is run entirely local. None of those tokens go to the cloud."
What models run on your Mac?
At launch, users can pick from three local models.
| Model | Developer | Notes |
|---|---|---|
| Gemma E4B | Smaller, faster option | |
| Qwen3.6 35B-A3B | Alibaba | Larger, more capable |
| Qwen3.6 35B (Perplexity-tuned) | Perplexity / Alibaba | Recommended by Perplexity |
Two of the three are based on Qwen, a model family developed by Chinese tech company Alibaba. When VentureBeat asked whether enterprise customers had raised concerns about running a Chinese-developed model on company machines, Staff argued the risk is neutralised when the model runs locally: the data never leaves your computer, and open-weight models, meaning models whose underlying code is publicly available for inspection, can be examined by anyone. MacOS also constrains what the agent can do through its built-in security framework.
That argument is reasonable for data privacy. Whether government or regulated-industry customers accept it is a different question, and one Perplexity hasn't yet answered with independent audits. Our earlier look at Alibaba's Qwen3.8-27B on 21 August found the model family scoring level with mid-tier cloud rivals on coding benchmarks, which at least grounds the capability claim.
What does this mean in practice?
Perplexity showed three demos. A lawyer updated a confidential draft brief while a cloud agent pulled public case law. A finance analyst ran a roughly 40-minute background task combining confidential projections with public market data. A small-business owner kicked off a marketing analysis from her phone while the agent reached her Mac at the office to process local customer data.
All three are jobs people currently do manually or refuse to give to an AI because the data's too sensitive. That's the real pitch here: not speed, but permission.
On cost: tokens processed locally use no credits. You pay credits only for the cloud orchestration piece.
One open question is consumer data use. Perplexity says it isn't using this data for training, but hasn't published the detail for non-enterprise accounts. Enterprise contracts can include zero-data-retention terms. For Pro or Max subscribers with privacy concerns, Perplexity's incognito mode and its existing opt-out toggle are the best available tools until the company publishes clearer terms.
Common questions
Does this work on Windows or older Macs?
No. Hybrid compute requires an Apple silicon Mac, the chip design Apple introduced in late 2020, running macOS 15 or later. Windows and Intel-based Macs aren't supported.
Who can use it right now?
Pro subscribers, Max subscribers and enterprise customers can opt in through the desktop app. A free-tier rollout hasn't been announced.
Is the Alibaba-made model a security risk?
Perplexity's position is no: the model runs entirely on your device, its code is publicly available for inspection, and macOS limits what it can access. Independent security researchers haven't yet published assessments of this specific setup.
