Perplexity's AI agents now shift work between Mac and cloud based on privacy needs

Perplexity has introduced Hybrid Compute, a feature that intelligently divides AI agent workloads between Apple's local hardware and cloud infrastructure. The system detects when a task involves sensitive information and routes that portion to smaller models running directly on your Mac, while keeping the rest of the job in the cloud.

How the privacy detection works#

At the heart of Hybrid Compute sits a local Privacy Gate—a Perplexity-trained filter that scans for personally identifiable data like names, addresses, account numbers, and credentials. When it identifies something sensitive, users can decide whether to process that step locally on their device instead. Before work begins, they review the system's flagging decisions and can override them if needed.

When it's time to start, users can pick from options including Gemma 2B, Qwen3.6 35B-A3B, or Perplexity's own fine-tuned version of Qwen. The company handles model setup through its desktop app, eliminating the need to open terminals or manage installations manually.

Economic and capability tradeoffs#

Running computations locally eliminates token charges—a meaningful cost savings when inference fees are shared across the industry. However, there's a performance trade-off: Perplexity operates its most advanced models remotely while relying on smaller, less capable ones for local work. Users control this balance depending on their privacy and cost priorities.

The system maintains task continuity across cloud and device boundaries. Agents can transition a step from the cloud to a Mac without restarting, and the local portion feeds back into the broader workflow without exposing the private data it handled.

Enterprise controls and cross-device handoff#

For organizational users, Perplexity provides device-level governance—company-wide rules governing what stays local and audit trails showing what data leaves each machine. The capability also supports handoffs between devices: users can begin a task on iPhone and transfer local work to their Mac without starting fresh.

This complements a similar feature Perplexity rolled out for Nvidia's DGX Spark, though with reversed logic. DGX Spark starts locally and contacts cloud models only with permission, whereas the Mac version defaults to the cloud and pulls work inward as needed.

Hardware requirements