Perplexity Computer Now Runs Fully on Your Mac

Perplexity Pro, Max and Enterprise subscribers can now run the full Computer experience locally on their own Mac, including the harness and orchestrator. Local availability opened Oct. 20. Engadget

Installation follows a fixed path. Users download the latest Perplexity Mac app and follow the in-app instructions to install a local model. The requirement is a recent Apple Silicon Mac with at least 36GB of RAM.

Computer does not operate as a thin client tied to a hosted agent. The orchestration logic, the code that plans tasks, hands work to smaller sub-agents and manages tool calls, runs on the device itself, along with model inference, the heavy calculation where the model produces answers. Readers who know the usual split between the control plane, the system that directs work, and inference will see the shift. Perplexity puts both on the Mac.

Local users pick one of two models: Qwen 3.6 or PPLX Pro. PPLX Pro is Perplexity's version of Qwen 3.6 with extra post-training, added tuning after initial training, for Computer use. The difference is structural. The base open-weight model, a publicly shared model anyone can run, covers general reasoning, while the tuned variant is set for Computer's specific command format, tool definitions and multi-step planning loops.

That local core can still draw on cloud models. Perplexity introduced Hybrid Compute in early September to split work between frontier models, large commercial models run in the cloud, and local models. Computer users can allow the software to call on more than 15 frontier models when needed. Any code or tool use a task requires runs in an isolated sandbox, a sealed-off area that keeps that execution separate from the rest of the system.

This builds on earlier work. Perplexity has pointed to Mac unified memory, memory shared between processor and graphics, for fast local AI and agent processing. Perplexity Its Personal Computer concept centers on a dedicated Mac mini that can run 24/7 and connects to local apps and Perplexity's secure servers. That Personal Computer became available to everyone on Mac in May, with the ability to work with local files and operate on the web. TechCrunch

The broader context here is the cost of hosted computer-use agents. They are easy to demonstrate and expensive to run at scale, with questions around credential scope, data leaving the device and delays from repeated tool calls. By moving the harness, orchestrator and execution sandbox onto user hardware, that tradeoff flips. Inference and repeated steps stay local and close to memory, while frontier models become a burst resource rather than the default path.

In my view, the two limits that will shape real deployments are memory capacity and routing policy. Thirty-six gigabytes excludes most base laptops and points local Computer at higher-spec desktops and MacBook Pro-class machines, where unified memory bandwidth can supply both the orchestrator and the active model without paging to slower storage. That is a reasonable constraint. Agent workflows break in confusing ways when context, tool outputs and intermediate plans compete for scarce memory.

Worth flagging for enterprise users is the permission model for Hybrid Compute. Local-first does not mean disconnected. Once a user allows fallback to frontier models, the prompts, file excerpts and browser state needed for that step leave the device. The sandbox still contains code execution, but data control now depends on routing choices made for each task. Administrators will want clear rules for when local inference is enough and when escalation is allowed, especially for workflows that touch local files.

Looking further out, what this enables is more practical personal automation. A Mac mini running around the clock with local orchestration, able to use local apps and files, browse when told to, and send only selected reasoning steps to larger models, looks less like a chatbot with tools and more like steady infrastructure. That is a quieter goal than full autonomy for everything, and more useful because of it.