Perplexity is expanding its Mac-based AI offering with an option to run the full Computer experience on a user’s own machine rather than relying on cloud processing. The catch is substantial but clear: the feature requires a recent Apple Silicon Mac with at least 36GB of RAM, and it is initially available to Perplexity Pro, Max and Enterprise subscribers.

The change goes beyond putting a single language model on a laptop. Perplexity says the local setup includes Computer’s harness and orchestrator, the components that manage an AI task and coordinate the work needed to carry it out. In practical terms, that means an eligible Mac can host the surrounding workflow system as well as the model selected to perform the task.

For people who use AI around sensitive personal or work material, the headline is straightforward: some tasks can remain on the Mac. Perplexity’s example is a request to make a prioritized to-do list using information in Notes and Mail. With the local configuration, the company says the email data involved in that workflow does not need to be sent to the cloud.

What has changed from Hybrid Compute?

The company had previously introduced Hybrid Compute for its Computer app at the beginning of September. That mode let users divide AI workloads between local models on their computer and more powerful remote, or “frontier,” models.

The new local mode takes that idea a step further. Rather than treating the Mac as one participant in a split workflow, it allows the whole Computer environment to run on the device, provided the hardware meets the memory requirement. Hybrid operation is still relevant, however, because users can authorize the app to call on remote models if they decide a specific task would benefit from additional help.

This distinction matters. “Local AI” can describe several very different setups: a local model that merely drafts text, a local assistant whose requests are still managed remotely, or a complete local task environment. Perplexity is positioning this release in the third category by saying the harness and orchestrator can stay on the Mac along with the chosen local model.

A quick guide to the technical language

  • Local model: an AI model installed and run on the user’s own computer instead of a remote server.
  • Orchestrator: the software layer that coordinates stages of a larger AI-driven task.
  • Harness: the surrounding system used to run and manage the model’s task workflow.
  • Frontier model: a high-end remote AI system that can be brought in for work a local setup may not be suited to handle.
  • Sandbox: an isolated environment intended to separate code or tool execution from the rest of a system.

Two local model choices, with a cloud fallback

Computer’s local mode offers two models: Qwen 3.6 and PPLX Pro. PPLX Pro is Perplexity’s own offshoot of Qwen 3.6, with additional post-training intended to tune it for use with Computer.

Post-training is the refinement work done after a foundational model has been trained. In this case, Perplexity’s stated goal is not simply to provide a general-purpose local model, but one better prepared for Computer’s workflow-oriented use. That framing is important because a model’s raw capability is only one part of an automated task system. It also needs to interact predictably with the software that gives it tasks, manages tools and handles execution.

Users are not locked into local processing for every request. Perplexity says they may authorize Computer to use one of more than 15 models when a task would benefit from a frontier system. The word authorize is meaningful here: the cloud handoff is presented as a user-approved choice, not as an invisible requirement of every local workflow.

That flexibility creates a practical decision point. A task involving private Notes or Mail content may be a strong candidate for local processing when keeping that material on-device is the priority. A different task may call for a remote model, but doing so changes the privacy boundary. Running the core Computer experience locally does not mean every future request remains local automatically if the user chooses to involve a frontier model.

Privacy gains are real, but the boundary still matters

The strongest appeal of a complete on-device setup is control over where sensitive information is processed. Perplexity specifically points to data from Notes and Mail as an example of material that can remain off the cloud in a local workflow. For users deciding whether AI assistants belong near their personal correspondence or work notes, that is a materially different proposition from sending the same content to a remote service.

There is also a spending angle. Perplexity says local execution gives users more control over credit spend. The logic is simple: work handled by the machine a person already owns does not have to consume cloud resources in the same way as work performed by remote systems. The company is pitching local capability as a route to both privacy and greater control over how AI usage is allocated.

Still, local processing should not be reduced to a blanket promise that every aspect of every AI job is private. The relevant question is which model and workflow path a user has chosen for the particular task. If a user authorizes a frontier model, that task is no longer solely an on-device operation. A thoughtful workflow means treating local mode as a useful control, then paying attention whenever a cloud option is selected.

The practical benefit is not that a computer becomes magically disconnected from outside services; it is that the user has a clearer opportunity to keep eligible work on the device when that is the preferred trade-off.

Why 36GB of RAM is a meaningful requirement

The 36GB memory threshold is not a small footnote. It sharply defines who can use the full local experience. RAM is the fast working memory a computer uses while active programs and data are in use. Running an AI model locally can demand a significant amount of it, and hosting the wider Computer workflow adds further requirements beyond ordinary everyday apps.

For potential users, the takeaway is to check the exact memory configuration of their Mac before treating local Computer as an available option. Owning an Apple Silicon Mac on its own is not enough; it must be a recent model with at least 36GB of RAM. The requirement may make the feature most attractive to people who already have suitably equipped hardware, rather than people looking for a reason to buy a new machine.

That limitation also explains why Hybrid Compute remains important. A split approach can give users some local participation without insisting that every piece of the workload run on their machine. The all-local option instead prioritizes maximum on-device control for systems capable of carrying it.

Tool use runs inside an isolated sandbox

Perplexity says any code execution or tool use required by a task happens in an isolated sandbox environment. A sandbox is a contained workspace designed to limit how that execution interacts with the broader system. In an AI workflow, this matters because the assistant may need to do more than generate prose; it may need to invoke a tool or run code as part of completing a task.

The company also says the same security protocols apply to the cloud version of Perplexity Computer. That does not make local and cloud execution identical from a data-location perspective. It does indicate that sandboxing is presented as a consistent protection around code and tool activity, rather than a special measure reserved only for local mode.

For subscribers, the rollout process is simple in outline. Install the latest Perplexity Mac app, then follow its in-app instructions to download and install a local model. Eligibility is limited to Pro, Max and Enterprise plans, along with the stated Apple Silicon and memory requirements.

Local AI is often discussed as though it is solely a performance contest between a small model on a personal machine and a large model in a data center. This release points to a broader question: where should the task manager, model, tools and the data used by an assistant reside? Perplexity’s answer for qualified Macs is that they can all reside locally, while leaving a deliberately optional route to remote models for tasks where users decide the trade-off is worthwhile.

That choice-oriented approach will be the key practical feature. Subscribers with the right Mac can keep suitable workflows self-contained, use Qwen 3.6 or Perplexity’s PPLX Pro model locally, and make a conscious decision if they want to extend a job to one of the available frontier models. For more perspective on getting the most from existing Apple hardware, see how to make an older iPhone feel faster before replacing it.