Your lawyer probably shouldn’t be uploading client files to a cloud AI server. Perplexity knows this, and its new Hybrid Compute feature is a direct response to that exact problem. As reported by Engadget, the company has added a way to split AI tasks between a powerful cloud model and a local one running entirely on your Mac, so the sensitive stuff never leaves your machine.
Here’s the core idea: you’re working on something that needs serious AI horsepower, like drafting a legal brief or analyzing confidential documents, but you don’t want private information sent to OpenAI’s servers or anywhere else. Hybrid Compute routes the sensitive parts of the task to a local model on your computer, while the heavier lifting goes to a frontier model like GPT-5.6 Sol or Opus 5 in the cloud. You get the best of both without sacrificing privacy on the parts that matter.
Before any of this kicks in, Perplexity’s app runs a new privacy classifier that automatically flags files or data it thinks should stay local. You get to review those suggestions before anything happens, which is a smart touch. You’re also in control of which local model handles the private side of things. Options right now include Gemma E4B and two versions of Qwen’s 35-billion parameter 3.6 model, one of which Perplexity post-trained itself. Installing any of these doesn’t require touching your Mac’s terminal, which is a genuine win for anyone who doesn’t want to mess with command lines.
While the task runs, you’ll see a live view of your CPU, GPU, and memory usage, plus a sidebar tracking token consumption. And here’s a nice detail: tokens processed by the local model are free. You only pay for what hits the cloud. So beyond privacy, there’s a real cost argument here for people doing high-volume work who don’t always need the most powerful model available.
Perplexity’s Jon Staff, who leads the company’s Mac products, was pretty upfront about the trade-offs. A fully cloud-based setup will almost always produce better raw output. But not everyone needs the absolute best model for every job. For some users, keeping data private and keeping costs down matters more than squeezing out the last bit of quality. That’s a fair and honest position, and it’s good to see the company say it plainly rather than oversell the feature.
This fits into a broader trend of AI companies trying to give users more control over where their data goes. On-device AI has been a talking point for a while now, but most implementations have been limited or underwhelming. Hybrid Compute is interesting because it’s not trying to replace cloud AI entirely. It’s using local processing strategically, only where it’s actually needed.
There are some real limitations to know about before you get excited. Hybrid Compute is Mac-only for now, requires macOS 15, runs on Apple Silicon, and Perplexity recommends at least 32GB of unified memory. That’s a fairly specific hardware requirement that rules out a lot of machines. It’s also limited to Pro, Max, and enterprise subscribers.
- Splits tasks between cloud models (GPT-5.6 Sol, Opus 5) and local models
- Local model options: Gemma E4B and two Qwen 3.6 35B variants
- Built-in privacy classifier flags sensitive files automatically
- No terminal required to install local models
- Local model tokens are free, no charge for on-device processing
- Requires Apple Silicon Mac with macOS 15 and 32GB unified memory
- Available to Pro, Max, and enterprise subscribers
So this is a genuinely useful feature for a specific type of user, particularly professionals who work with confidential data and want AI assistance without the compliance headache. It’s not a must-have for everyone, and the hardware bar is high. But if you’re in that target group and you’re already on Perplexity’s higher tiers, this is worth paying attention to.