- Published on
Perplexity Releases Hybrid Compute on Mac: Cloud Agents Orchestrate Down to a Local Model, Gated On Device
- Authors
- Name
- DeepHorus
Perplexity has unveiled a new hybrid compute system for its Mac application, designed to address a core dilemma in AI assistants: balancing powerful cloud-based processing with the need to protect sensitive user data. The feature, now live for Pro, Max, and Enterprise subscribers, splits AI tasks between frontier models in the cloud and a compact local model running directly on Apple Silicon Macs.
The system inverts a common approach by starting every task in the cloud, where models handle complex web search and reasoning. When a step requires accessing private files or data, it is seamlessly handed off to the local model on the user's Mac without losing context. This ensures sensitive information like client records or financial documents never leaves the device.
The critical component is an on-device privacy gate. Before any data is sent to the cloud, a newly open-sourced classifier called PII-Tracer inspects it. This 0.6 billion parameter model applies one of four actions: keeping data local, masking sensitive spans, refusing the action, or asking for user consent. It specifically targets credentials, payment details, and government IDs.
Perplexity trained PII-Tracer on roughly 714,000 samples. Performance results show it achieves the highest character F1 score of 0.629 among tested detectors. Its consistency is particularly notable, finding every mention for 79.4% of recurring identifiers, significantly outperforming models like GPT-5.6-sol.
The hybrid compute feature is available on any Apple Silicon Mac with macOS 15 or later and at least 24GB of unified memory. The local model installs with one click and requires no separate runtime or API keys. Local processing consumes no cloud credits, offering a privacy-focused alternative for enterprise and power users.
At launch, three local models are supported, including Gemma 4 E4B and Qwen3.6. The system also integrates with iPhone, allowing a task to be triggered remotely while sensitive steps execute on a desk-bound Mac. This positions an always-on Mac mini as a potential dedicated inference node.
The launch responds to a fundamental user need: leveraging AI's full capability without compromising security. By giving users control over what data crosses the boundary, Perplexity aims to enable agentic assistants in professional settings where privacy is non-negotiable.
This move signals a growing industry focus on hybrid architectures. As AI assistants become more capable, the ability to securely handle sensitive workflows will likely become a standard requirement rather than a premium feature.
