Perplexity launches Hybrid Compute to keep sensitive data off the cloud

Perplexity on Tuesday launched Hybrid Compute, a feature for its Mac app that splits AI agent tasks between cloud-based frontier models and a local model running on Apple silicon, routing sensitive data to the user's machine so it never leaves the device.

The system works like a dispatcher: a frontier cloud model breaks a task into subtasks, handling web research and heavy reasoning remotely, while anything touching private files or personally identifiable information gets delegated to a local model on the Mac. An on-device "Privacy Gate" — a classifier Perplexity trained in-house — scans for names, addresses, account numbers, and other sensitive content before anything is transmitted to the cloud.engadget+2

"By combining these two together, we can get that maximum intelligence from the frontier models, but we also get the security and the privacy that comes with local," Jon Staff, who leads Perplexity's macOS and iOS engineering teams, said during a press briefing reported by VentureBeat.venturebeat

How It Works

Before Hybrid Compute begins delegating work, users can review what the Privacy Gate has flagged and decide what stays on their machine. At launch, local model options include Google Alphabet Inc.'s ↘ 1.62% Gemma E4B, Alibaba's ↘ 3.87% Qwen3.6 35B, and a Perplexity post-trained variant of Qwen3.6 — the company's recommended choice. Installation requires a single click inside the app with no terminal setup.appleworld+1

Tokens generated locally cost nothing; credits are consumed only for cloud orchestration. Users can also start a task from an iPhone and send its local steps to their Mac remotely.9to5mac+2

Perplexity has open-sourced the PII classifier, developed with its Secure Intelligence Institute, along with a new evaluation standard for PII detectors.appleinsider+1

Use Cases and Limitations

Perplexity demonstrated the feature with a lawyer updating a brief against privileged case files while a cloud agent pulled public case law, and a private equity associate reworking a financial model against confidential projections — a task that ran roughly 40 minutes with no human input.venturebeat

Staff acknowledged trade-offs. "A fully frontier output is going to almost always be better in terms of raw artifact creation. It's more expensive, it's more capable," he said, but added that for many users, data privacy and cost savings outweigh peak performance.engadget

Availability

Hybrid Compute is available today for Pro, Max, and enterprise subscribers on Apple ↗ 1.32% silicon Macs running macOS 15 or later. Perplexity lists 24GB of unified memory as the minimum, with 32GB recommended for the best results. Enterprise admins can set organization-wide sensitivity policies and audit what leaves each device. Windows and Linux support will follow later.9to5mac+2