Perplexity Computer adds local execution on Macs with PPLX Pro

• Perplexity Computer can run its model and agent entirely on a Mac, including without Wi-Fi, according to the launch thread.
• PPLX Pro is a post-trained version of Qwen 3.8, per Perplexity's model announcement.
• Fully local tasks consume no Computer credits, but access requires a Pro, Max, or Enterprise subscription, as the launch coverage notes.
• Fully local execution carries a recommendation of 36 GB unified memory on supported Apple silicon Macs, according to the availability report.

The product page describes a Privacy Gate that masks personal information before cloud processing and restores real values in the result. The launch post also lets the local orchestrator call on 15+ frontier models when authorized.

Perplexity's launch history separates three stages of its Mac rollout:

1. May: Personal Computer gained access to local files and native Mac apps.
2. September: Hybrid Compute coordinated cloud reasoning with local models processing private information.
3. October 20: The model, harness, and task execution can all run on-device, keeping planning, reasoning, and file processing on the Mac.

The launch post and Perplexity's model announcement identify two downloadable local options:

• PPLX Pro: A post-trained Qwen 3.8 model. Perplexity specified 27 billion dense parameters in a follow-up.
• Qwen 3.6: A separate local model choice offered alongside PPLX Pro.

Native-app access is the best part of this launch: the demo pairs local inference with actions across files and apps.

Computer can carry out a budget update across several apps, according to Perplexity's workflow example:

1. Read a project's to-do list in Notes and a teammate's budget request in Messages.
2. Pull planned and actual spending from Excel.
3. Prepare a status update with open tasks and remaining budget, then send a meeting invitation.

Code and tool execution run in isolated sandboxes with controlled access to files and connected apps, according to the security description. Perplexity says the cloud version uses the same security protocols.

The broader Personal Computer setup guide lists three macOS permissions:

• Accessibility: Allows the agent to click and type.
• Screen & System Audio Recording: Allows it to see what's on screen.
• Full Disk Access: Allows it to read files.

Hybrid starts the task in the cloud, with cloud models handling research and reasoning while the Mac processes private files, according to the product documentation.

Privacy Gate checks personal information on the Mac and can take three paths:

• Keep content local.
• Mask it before cloud use.
• Ask the user to approve sharing.

Locally started tasks can also escalate to cloud services after authorization. The launch post names current information, browser use, and connected apps as reasons for that escalation.

Availability covers Pro, Max, and Enterprise subscribers on supported Apple silicon Macs, per Perplexity's availability announcement. The setup instructions require the latest Mac app and an in-app local model download.

The 32 GB figure quoted in a community reply matches Hybrid's recommendation on the product page, which distinguishes the two modes:

• Fully local: 36 GB unified memory recommended.
• Hybrid: 32 GB unified memory recommended.

Perplexity's product page draws two additional network boundaries:

• Messages and invitations: Sending a message or meeting invitation still uses the relevant app's service, even when the AI runs fully locally.
• Remote initiation: In the iPhone example, the task starts in the cloud while the Mac processes private files locally. The resulting deck is accessible from other devices.