Perplexity is giving its AI agents a new way to work with sensitive information on the Mac. The company has launched Hybrid Compute, a system that splits individual Perplexity Computer tasks between cloud-based AI models and a local model running on the user’s Apple silicon Mac.

The idea is fairly simple, but potentially important: the cloud handles the parts of a task that benefit from powerful frontier models, while the Mac handles sensitive files, personal information, and other work that users may not want sent to a remote server. Perplexity says the two sides can work together within the same task, so users don’t have to manually separate private information from cloud-based prompts.

Perplexity introduced Personal Computer for Mac earlier this year, giving Computer access to local files, applications and other resources on a Mac. Hybrid Compute takes that local capability a step further by adding a local AI model to the workflow.

When a task starts, Computer can use cloud models for research, web searches, planning and complex reasoning. When it encounters sensitive information stored on the Mac, the workflow can hand that portion of the task to the local model instead.

A privacy gate running on the Mac controls what information can leave the device. Depending on the situation, it can mask sensitive information, keep it entirely local, refuse an action or ask the user for permission before allowing information to reach the cloud.

Perplexity says the privacy gate uses an on-device classifier to identify personally identifiable information such as names, addresses, account numbers and secrets. Credentials, payment-card numbers and government identification numbers receive stronger protections, with the system able to keep them local or rewrite a request so that the cloud model can continue without exposing the protected information. The company has also open-sourced the PII classifier used for this purpose.

That means Hybrid Compute isn’t simply a choice between “local AI” and “cloud AI.” Instead, Perplexity is trying to make both approaches part of one agentic workflow.

For example, an investment team could ask Computer to research a prospective company using public filings, market information and comparable transactions in the cloud. The local model could then examine confidential deal documents stored on the Mac and compare their assumptions with the public research without sending those documents to Perplexity’s cloud.

The same approach could be used by law firms, advertising agencies and other organizations handling confidential information. Perplexity gives the example of lawyers researching case law online while keeping privileged documents on the Mac. The local model can extract relevant facts and turn them into anonymized research questions that the cloud model can use.

The Mac therefore becomes more than a place where files happen to live. It becomes another compute environment inside the AI agent’s workflow.

Three local models are available

Perplexity is launching Hybrid Compute with three local model options: Gemma 4 E4B, Qwen3.6 35B-A3B and a Perplexity model. Users can download a model directly from the Mac app without installing Ollama or manually configuring a local inference runtime.

Once a local model is installed, users can open Perplexity’s model selector, choose Hybrid and select the local and cloud models they want to use for the task.

The local processing also doesn’t consume cloud credits for the work handled by the local model, according to Perplexity. No API key or separate local runtime setup is required.

There are some hardware requirements, though. Hybrid Compute requires an Apple silicon Mac running macOS 15 or later with at least 24GB of unified memory. Perplexity recommends 32GB for the best results.

That memory requirement makes sense given the size of some of the models involved. Qwen3.6-35B-A3B, for example, has 35 billion parameters but uses a sparse architecture that activates roughly 3 billion parameters per token.

Perplexity has also been working specifically on making local inference fast on Apple silicon. Its Lily inference engine is designed around Qwen3.6-35B-A3B and Apple’s hardware, with separate optimizations for prompt processing and token generation. In Perplexity’s testing on an M5 Max MacBook Pro with 128GB of unified memory, Lily delivered an average 1.23x higher prefill throughput and 1.35x higher decode throughput than MLX-LM across the tested workloads.

That optimization work matters because Hybrid Compute needs local processing to happen quickly enough that switching between cloud and on-device models doesn’t make the overall agent feel sluggish.

Your iPhone can trigger the Mac

One of the more interesting parts of the system is that the Mac doesn’t necessarily need to be sitting in front of you.

Perplexity says Computer can be triggered from an iPhone while the Mac performs the local portions of the task. The Mac can access its local files and run the sensitive steps on-device, allowing users to start work remotely while retaining local processing for information that shouldn’t leave the computer.

Perplexity even suggests using a dedicated Mac mini as an always-on Computer machine. Because the Mac mini can remain powered on and connected, it can act as a local AI workstation that users control remotely from an iPhone.

That is a particularly interesting use of Apple’s desktop Macs. Instead of buying a Mac mini simply as a conventional desktop computer, users could potentially leave one running as a private local AI machine that handles sensitive parts of automated workflows.

Perplexity is betting on hybrid AI

Hybrid Compute arrives as the AI industry continues to move away from the idea that every AI workload has to run entirely in the cloud.

Fully local AI offers obvious privacy and cost advantages, but smaller local models don’t always match the reasoning and tool-use capabilities of the largest cloud models. Cloud AI, meanwhile, offers substantially more compute but requires users to trust a remote service with the information being processed.

Perplexity’s approach is to avoid forcing users to choose.

The company wants Computer to decide which parts of a workflow belong on the Mac and which belong in the cloud, with the privacy gate enforcing boundaries around sensitive information. That could make agentic AI more practical for professional workflows where confidential documents and powerful cloud-based reasoning need to coexist.

Hybrid Compute is now available to Perplexity Pro, Max and Enterprise subscribers on Apple silicon Macs running macOS 15 or later. Enterprise administrators can also establish organization-wide policies controlling what information must remain on the Mac, what can be masked before being sent to the cloud, and which actions require user approval. Administrators can audit when information leaves a device as well.

For users with a compatible Mac, getting started is straightforward: install the latest Perplexity Mac app, download one of the supported local models, then select Hybrid in the model selector.

It’s an intriguing direction for Perplexity Computer. Rather than treating local AI as a replacement for cloud AI, Hybrid Compute turns the Mac into a private extension of the cloud-based agent—and, for sensitive workflows, that distinction could be the whole point.