How Perplexity Hybrid Compute Splits AI Work Between the Cloud and Your Mac

Perplexity’s Hybrid Compute is designed to give its Computer agent access to both kinds of AI processing: powerful frontier models in the cloud and smaller models running directly on an Apple-silicon Mac. The cloud handles web research, planning, and difficult reasoning, while the Mac performs steps involving private files, sensitive information, or local actions. Computer then combines the results into one workflow. 1

The key idea is not to run the entire task locally. Instead, Hybrid Compute keeps the cloud’s capabilities for work that benefits from remote infrastructure and moves sensitive steps to the user’s Mac.

How Hybrid Compute works

A task can begin in the cloud, where Perplexity’s Computer agent handles search, orchestration, and reasoning. When a step involves protected information or files on the Mac, the agent can delegate that part to a local model. The local result is returned to the workflow without sending the sensitive material to Perplexity’s servers, according to the company. 1 16

This division makes Hybrid Compute different from a fully local chatbot. Users retain access to cloud-connected research while using on-device inference for confidential work.

What the Privacy Gate does

The local Privacy Gate is the control point between the Mac and cloud processing. Perplexity says its on-device PII classifier examines task information before it leaves the computer. Reported coverage includes prompts, tool outputs, memory, and logs, as well as files and other task content. 1 5

The classifier is designed to identify information such as:

• names and addresses;
• account numbers and credentials;
• payment-card details;
• government identification information; and
• other sensitive secrets or personal data. 1 3

When the gate detects potentially sensitive information, users can review the flagged items and choose how to proceed. Available actions include keeping the information on the Mac, allowing it to be uploaded, masking sensitive values before cloud processing, refusing the action, or requiring explicit approval before escalation. 1 3

With masking, sensitive values are replaced by stand-ins for the cloud step and restored in the returned answer, according to Perplexity’s description. 1 This can let a cloud model work with the structure of a task without receiving the original identifiers.

The Privacy Gate reduces the chance of accidental disclosure, but it is not a guarantee that every confidential fact will be identified. The privacy outcome still depends on the classifier, the organization’s routing rules, and the user’s approval decisions.

Which local models are available?

The initial local choices include Gemma 4 E4B, Qwen3.6 35B-A3B, and a Perplexity model post-trained for Computer. 1 4

These compact local models are not intended to replace the strongest cloud models for every task. Their role is to process private context and perform local actions while the cloud models provide broader research, planning, and frontier-level reasoning.

How to install and enable it

Perplexity’s setup is designed to avoid a separate local-model stack. In the Mac app, users download a local model in one click, select Hybrid in the model picker, and choose the local and cloud models they want to use. The supplied product information says no Ollama installation, manual runtime configuration, or API key is required. 1

Local tokens do not use cloud credits

Inference handled by the local model does not consume the user’s cloud credits. Cloud portions of the same task still use cloud inference and remain subject to the applicable cloud allocation. 1 2

That makes the feature relevant for both privacy and cost control: sensitive processing can remain on the user’s hardware, while local token generation does not reduce the cloud-model allowance. The trade-off is that local speed and output quality depend on the Mac’s unified memory, the selected model, and the complexity of the step.

Starting a task from an iPhone

Hybrid Compute can also work across Apple devices. A user can start a Computer task from an iPhone; the task begins in the cloud, then the Mac accesses local files and performs the sensitive steps. The completed result can be viewed across devices. 1 5

For remote access to local inference, Perplexity suggests using a dedicated, always-on Mac mini. 1 That setup is more convenient than manually moving files between devices, but it also means the Mac must be available when the local step is needed.

What Perplexity has open-sourced

Perplexity says it has open-sourced the on-device PII classifier used by the Privacy Gate. The company describes the classifier as having been developed with its Secure Intelligence Institute. 1 6

Making the classifier available for inspection can give security and enterprise teams more visibility into the mechanism that triggers local routing. It does not, by itself, prove that every category of confidential information will be detected in every context.

Example workflows for professionals

Legal work

In Perplexity’s example, the cloud model researches case law while the local model summarizes privileged case files, extracts facts, creates anonymized research questions, and updates a local draft brief. The intended benefit is that client data and case files remain on the Mac while public research uses cloud models. 1

Private equity and investment research

For an investment team, the cloud can gather public filings, comparable transactions, and market information. The local model can read confidential deal materials, extract terms and assumptions, compare them with public research, and update an investment-committee deck or LBO analysis. 1

These examples illustrate the division of labor: public information and broad research go to the cloud, while confidential source material stays within the local processing boundary.

Main trade-offs

Privacy: Sensitive files can remain on the device, but protection depends on accurate classification and careful approval settings. 1 3

Cost: Local inference does not use cloud credits, while cloud reasoning and research continue to consume cloud resources. 1 2

Speed: Cloud models offer fast access to large-scale inference, while local execution avoids uploading protected files. Local performance varies with the Mac’s memory and model choice. 1

Capability: Frontier cloud models remain better suited to demanding reasoning, web research, and planning. Local models trade some capability for on-device processing and greater data control. 1

Requirements and availability

The feature requires an Apple-silicon Mac running macOS 15 or later and at least 24 GB of unified memory. Perplexity recommends 32 GB for the best results. 1 3

The supplied product information lists Perplexity Pro, Max, and Enterprise subscribers as eligible users. 1 Reporting also describes an Enterprise opt-in path. 11

Enterprise administrators can define organization-wide policies for data that must remain local, information that may be masked before cloud processing, and actions that require user approval. They can also audit when information leaves a device. 1

Is there a Windows or Linux version?

The available product materials describe Hybrid Compute as a Mac and Apple-silicon feature. They do not provide a Windows or Linux release date, so there is no supported arrival timeline to report.

For now, Hybrid Compute’s appeal is its compromise: it does not ask users to choose between the capability of cloud AI and the control of local processing. Instead, it uses a local Privacy Gate to decide which parts of an agent workflow can stay on the Mac and which can be sent to the cloud.