Agentic assistants become useless when the data they need—deal files, privileged documents, client records—cannot leave the user’s device because of privacy rules. Sending that information to a cloud model risks exposure, while keeping everything on‑device sacrifices the power of frontier models for reasoning and web search.
Perplexity’s hybrid compute solves this split‑brain problem by starting every task in the cloud where a large model handles search, planning and long‑horizon reasoning. When a step touches private data, the task is handed off to a compact model running locally on the user’s Mac, without losing context or restarting the workflow. The two halves are merged into a single answer, giving the user both cloud‑scale intelligence and on‑device confidentiality.
The critical piece is the on‑device privacy gate. Before any snippet of a protected file reaches the cloud, a 0.6 B PII‑Tracer classifier inspects the text and decides—keep it local, mask the sensitive spans, block the action, or ask the user for consent. Masked values are replaced with stand‑ins for the cloud step and restored when the result returns, so the model never sees raw identifiers.
Hybrid compute is live today for Pro, Max and Enterprise subscribers on any Apple‑silicon Mac with macOS 15 or later and at least 24 GB of unified memory (32 GB recommended). The local model installs with one click from the Mac app, needs no Ollama, separate runtime or API key, and consumes no cloud credits for the on‑device work. Enterprise admins can set organization‑wide rules about what must stay on device, what may be masked, and what requires explicit approval, plus audit trails for any data that does leave the machine.
This approach lets legal, financial and healthcare teams use powerful AI assistants without breaking compliance, turning a structural limitation into a practical, deployable solution.
#AI #Product #HybridCompute #Privacy #MacOS #PII #DataSecurity #AIAgents