Perplexity Computer now runs fully on-device on Macs with 36GB+ RAM, using Qwen 3.6 or PPLX Pro

Perplexity has made it possible to run its full Computer experience, including the harness and orchestrator, entirely on a user's own Mac. The requirement is a recent Apple Silicon machine with at least 36GB of RAM. The feature is available starting today (Oct. 20, 2026) to Perplexity Pro, Max and Enterprise subscribers, Engadget reports.

What's new

In early September, Perplexity introduced Hybrid Compute, an update to its Computer app that split AI workloads between frontier models and local models running on the user's machine. Today's release goes further: the agent harness and orchestrator now run on-device too, so a task can run end to end without leaving the Mac.

Users can pick one of two local models:

• Qwen 3.6, the open model from Alibaba's Qwen team.
• PPLX Pro, a Perplexity offshoot of Qwen 3.6. According to Perplexity, the company did its own post-training on it to tune it for Computer use.

Parameter counts, context window and benchmark results for either local configuration were not disclosed in the announcement as reported.

Privacy and cost claims

Perplexity pitches the feature on data control and credit spend. Its example: with access to the Notes and Mail apps, Computer can generate a prioritized to-do list without any email data being sent to the cloud. The company says: "As more demanding workflows become practical to run locally, the machines people already own will take on a larger share of their AI work. This gives users more control over sensitive data and credit spend."

These are Perplexity's claims. No independent testing of local task quality, speed or privacy guarantees has been published.

Frontier escalation and sandboxing

Local mode is not strictly offline-only. If a task would benefit from a stronger model, users can authorize Computer to call one of more than 15 frontier models. Any code or tool execution a task requires runs in an isolated sandbox. Perplexity says these security protocols are the same ones used in the cloud version of Computer.

Availability and setup

• Plans: Pro, Max and Enterprise.
• Hardware: Apple Silicon Mac with at least 36GB of RAM.
• Setup: Install the latest Perplexity Mac app, then follow the in-app instructions to download and install a local model.
• Pricing: No separate fee for local mode was disclosed. Subscription pricing was not part of the announcement.

What this means

The 36GB RAM floor is the main constraint. It excludes base-configuration Macs and limits local mode to higher-memory machines, so the feature will reach only a subset of Perplexity's Mac users. That floor also indicates how much memory a capable agentic stack needs once the model, the orchestrator and the working context all sit on one device.

Strategically, Perplexity is building on an open-weight base (Qwen 3.6) and differentiating through post-training and its harness rather than pretraining. PPLX Pro follows that pattern. The open question is how large the quality gap is between local models and the frontier models Computer can escalate to. Perplexity has published no data on it, and that gap will determine how many tasks users actually keep on-device.

The privacy argument is real for tasks touching local mail and notes. But because the cloud-escalation option stays available, the actual data exposure depends on how often users authorize it and on what context gets passed along. Perplexity has not detailed that.