Perplexity Ships Hybrid Compute for Mac, Blending Cloud and Local AI

Perplexity released its hybrid compute feature for Mac this week, enabling a novel approach to AI agent architecture that addresses the inherent privacy challenges of processing sensitive user data. This new system splits AI tasks between powerful frontier models hosted in the cloud and a compact model running directly on the user's Mac. A critical component of this hybrid system is an on-device privacy gate, which is powered by an open-sourced classifier developed by Perplexity. This gate meticulously inspects data from protected files before any information is sent to the cloud, determining one of four actions: keeping the data local, masking sensitive portions, refusing the action entirely, or requesting explicit user consent. This ensures that confidential information, such as deal documents, privileged files, or client records, remains protected on the user's device. Hybrid compute is currently available to Perplexity Pro, Max, and Enterprise subscribers who are using an Apple silicon Mac equipped with macOS 15 or later and a minimum of 24GB of unified memory, with 32GB recommended for optimal performance. The local model can be installed with a single click from the Mac application, eliminating the need for external tools like Ollama, separate runtimes, or API keys, and crucially, local processing does not consume cloud credits. The orchestration of tasks within Perplexity Computer begins in the cloud, where advanced models handle complex operations like web searches, strategic planning, and long-horizon reasoning. However, when a task requires access to private or sensitive data, the system seamlessly hands off that specific step to the local model on the Mac. This handover occurs without interrupting the task's flow or losing contextual information, and the results from both cloud and local processing are then merged to produce a single, coherent output. This functionality represents an inversion of the local compute mode Perplexity introduced on NVIDIA DGX Spark the previous week, which initiated tasks on user hardware and escalated to cloud models with permission. The same orchestrator is employed, but the default operational direction is reversed. The system is designed to work with Perplexity's iPhone application, allowing tasks to be initiated remotely while sensitive computations are executed on a user's Mac, potentially acting as a dedicated local inference node, especially when configured as an always-on Mac mini. The privacy gate is described as the load-bearing component of this architecture, safeguarding user data by applying granular controls based on the classifier's analysis of the content and its sensitivity.