Skip to content
NewsProduct launch

Perplexity Hybrid Compute Routes Sensitive Files to a Local Model Before They Reach the Cloud

· by Pondero Newsdesk

The short version

Perplexity launched Hybrid Compute for its Mac app on September 1, 2026, splitting Computer agent tasks between cloud models and a local model on the Mac. A privacy gate built on an open-sourced PII classifier intercepts sensitive content before it leaves the device.

Perplexity Hybrid Compute Routes Sensitive Files to a Local Model Before They Reach the Cloud

An on-device PII classifier now intercepts agent tasks before they leave the Mac. Perplexity launched Hybrid Compute for its Mac app on September 1, 2026, per 9to5Mac, splitting Computer agent tasks between frontier cloud models and a local model on the device. Medical records, financial files, and similar sensitive documents are processed locally. Reasoning, web search, and planning continue in the cloud.

What shipped

Perplexity's Computer agent previously ran entirely in the cloud. Hybrid Compute changes the execution model. Each task starts in the cloud; when a step involves private files or sensitive data, the agent hands off to a local model running on the Mac.

The privacy gate is the core mechanism. An on-device PII classifier reads each task before it is sent to the cloud. Names, addresses, and account numbers are replaced with stand-in tokens, then restored when the answer returns. If a step is too sensitive to send even with masking, the classifier can process the content locally, refuse the action, or prompt the user for consent before any data leaves, per Perplexity. Perplexity open-sourced the PII classifier, developed in collaboration with its Secure Intelligence Institute.

Setup requires downloading PPLX Qwen 3.8 27B once, in one click, with no Ollama dependency and no API key, per the Perplexity product page as reported by 9to5Mac. Work processed locally consumes no cloud credits. Hardware requirements are Apple silicon with macOS 15 or later and at least 24 GB of unified memory; Perplexity recommends 32 GB for best results. Macs with 8 GB or 16 GB of RAM are not supported. Access is available to Perplexity Pro, Max, and Enterprise subscribers, per Perplexity.

Perplexity first signaled this direction in June 2026 when it announced plans to split agent tasks between local and cloud models. The September 1 release is the shipped implementation.

Why it matters

Privacy-gated AI agents have been discussed for years. This is among the first shipped consumer implementations that intercepts at the task level without requiring the user to choose between a local and cloud tool. The PII classifier decides automatically, with no configuration step. For operators running agents over regulated documents or client data, the distinction matters: compliance exposure comes from what reaches the cloud, not just what the model outputs.

The open-sourced PII classifier is worth tracking separately. Released via Perplexity's Secure Intelligence Institute, it could become a building block for teams that want to apply the same gate to other AI workflows on Apple silicon, outside Perplexity's own product stack.

The 24 GB memory floor creates a real deployment barrier. The M4 Mac mini base model ships with 16 GB; the M4 Pro Mac mini starts at 24 GB. Teams evaluating Hybrid Compute need to confirm their hardware before counting on it as a deployable option.

What to watch next

Perplexity has not announced a Windows path. Apple silicon is the only supported runtime, so Windows support would require a different NPU implementation. Whether the PII classifier becomes available as a standalone package for enterprise on-premise deployments is a separate open question. Both would extend this approach to a significantly larger audience than current Perplexity Mac subscribers.

Sources