# Perplexity Computer on Mac: Fully Local AI, Models and Requirements

Perplexity can now run the model and agent on an Apple silicon Mac, changing which tasks need a cloud connection.

By Paras Madan | 2026-10-21


Perplexity launched fully local Computer for Mac on October 20, 2026. The AI model, agent harness and task execution can now run on the device, rather than relying on the cloud to coordinate the work. It gives Mac users a way to complete supported tasks without uploading their private files.

The important change is where the agent's planning happens. A local model alone is only part of a private workflow if the system deciding what to do next still runs elsewhere.

## Perplexity Computer Mac requirements

The [official launch post](https://www.perplexity.ai/en-GB/hub/blog/perplexity-computer-now-runs-fully-local-on-mac) lists availability for Pro, Max and Enterprise subscribers on supported Apple silicon Macs. Perplexity recommends 36 GB or more of memory and offers Qwen 3.6 and PPLX Pro, its post-trained Qwen model, at launch.

To start, update the Perplexity Mac app, follow its instructions to download and configure a local model, then choose fully local execution.

The memory figure is a recommendation, not a published universal minimum. The announcement does not provide a complete device-by-device performance table or task benchmark. An Apple silicon badge alone does not tell you how comfortably a particular workflow will run.

## Fully local versus Hybrid Compute

Perplexity's earlier [Hybrid Compute](https://www.perplexity.ai/en-GB/hub/blog/introducing-hybrid-compute-on-mac) splits a task between cloud reasoning and on-device work. Its privacy gate can keep protected information local, mask it, refuse transmission or request consent before allowing it out.

| Mode | Where planning starts | Main tradeoff |
| --- | --- | --- |
| Hybrid Compute | Cloud, coordinating with a local model | Access to cloud reasoning while routing sensitive work locally |
| Fully local Computer | Mac, alongside the model and task execution | More work stays on device, with capabilities tied to local resources |

Hybrid remains useful when a task needs current public research alongside private documents. Fully local makes more sense when the necessary inputs already exist on the machine and the local model can handle the job.

**Local file processing and a fully local agent are different deployment choices.** The new mode moves the controller as well as the processing onto the Mac.

## Can Perplexity Computer work offline?

Yes, for work that can be completed with the local model and available local data. [9to5Mac's launch coverage](https://9to5mac.com/2026/10/20/perplexitys-computer-can-now-run-entirely-on-your-mac-even-offline/?extended-comments=1) describes offline operation and the shift from Hybrid Compute to running both model and agent on device.

Perplexity says users can authorize escalation for current information, browser use, connected apps or more advanced cloud reasoning. Supported local apps include Notes, Messages, Mail and Numbers, with permission.

Offline therefore describes the execution path for a suitable task. It does not make live web research possible without a connection, or remove the need to authorize external services when a workflow uses them.

## What to test before relying on it

A useful first test is a bounded task involving a small set of local documents, with an output you can inspect. Check whether the agent selects the right files, preserves figures correctly and asks before an action crosses the intended boundary.

Local execution reduces one category of exposure, but the agent still needs sensible file permissions and human review. Keeping an incorrect edit on your own machine does not make it harmless.

The launch also leaves detailed credit treatment and comparative performance insufficiently specified for a blanket cost claim. Subscription access is required; fully local should not be described as a free replacement for the product.

The practical opportunity is to reserve cloud calls for the parts of a workflow that genuinely need them. Whether that works depends on the task, the model and the memory available, rather than the word local alone.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn product architecture and meaningful launches into clear stories developers can use and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).