Do Son 
AI Agents Move From the Cloud to the Desktop
Generative AI is shifting from cloud Q&A toward agent workflows that operate desktop software. Power users now ask how to balance convenience, API spending, and personal privacy.
Perplexity has announced its answer. Mac users can now run AI workflows entirely on the device. The catch is memory: the Mac needs at least 36GB.
From Hybrid Compute to Fully Local
In early September, Perplexity launched Hybrid Compute. That feature lets users split AI workloads between cloud frontier models and local models. Now the company goes a step further.
Users can run the whole Perplexity Computer experience on their own Mac. This includes the control framework and the task orchestrator, which Perplexity describes as the most critical layers.
Hardware Requirements
A recent Apple Silicon Mac with at least 36GB of unified memory qualifies. With it, the full architecture can run offline or stay on the device while it performs AI tasks.
Why Go Fully On-Device
Perplexity says the update targets two pain points.
- Stronger data privacy: a user can let Computer read Notes and Mail, analyze email on the device, and build a prioritized to-do list. No sensitive email data travels to a cloud server.
- Lower API and credit spend: heavy users can hand routine automation to a Mac they already own. That sharply cuts cloud compute usage.
In the company’s words, as more demanding workflows become practical to run locally, the computers people already own will carry a larger share of AI work. That also gives users more control over sensitive data and spending.
Two Local Models and a Sandbox
Perplexity Pro, Max, and Enterprise subscribers with eligible hardware can try it now. They download the latest Perplexity Mac app and follow the in-app prompts to install one of two local models.
The Model Choices
- Qwen 3.6: a well-regarded, recent model from Alibaba’s open-source ecosystem.
- PPLX Pro: Perplexity’s own variant of Qwen 3.6. Post-training tunes it for computer operation and workflow tasks.
Flexibility and Safety
Users keep a path back to the cloud. If a complex task needs more power, they can authorize the app to call any of more than 15 cloud frontier models.
Safety measures stay in place as well. To keep an agent from running out of control, all code execution and tool calls take place in a strictly isolated sandbox. The security standards match those of the cloud version of Perplexity Computer.
A New Baseline for Mac Memory
On-device AI agents have entered practical use, and large unified memory is the ticket in. Many buyers of MacBook Pro and Mac Studio models once hesitated over upgrades to 36GB or 48GB. Perplexity now draws a clear line. To run the orchestrator and a model such as Qwen 3.6 or PPLX Pro smoothly, 36GB looks like the minimum.
At a Glance
| Category | Details |
|---|---|
| Hardware | Recent Apple Silicon Mac with at least 36GB of unified memory |
| Local models | Qwen 3.6, or PPLX Pro (tuned for computer operation) |
| Privacy and security | Data such as Mail and Notes stays on the Mac; code and tools run in an isolated sandbox |
| Availability | Perplexity Pro, Max, and Enterprise subscribers, through the latest Mac app |
| Cost | Less cloud credit spend, because the Mac carries the workload |
Support Our Threat Intelligence
Find our tech and OS security coverage helpful? Support our work today and unlock a 100% ad-free reading experience!
Buy Me a Coffee
