AIBRIEF
Hybrid Compute splits one task between a local LLM and a cloud model, but the privacy classifier is Perplexity's own.
Picture a lawyer building a brief against existing case law. The client data stays on the Mac; the reasoning about precedent goes to a cloud AI. That is the shape of Perplexity's new Hybrid Compute, a feature in the company's Mac app that splits one task between a frontier model like Claude or GPT and a smaller open-weight model running on the laptop.
A Perplexity-trained privacy classifier flags files and prompts the user should keep local. Before the task runs, the user reviews which files Hybrid Compute wants to gate and picks which model handles which part. Local options at launch include Google's Gemma E4B and two variants of Alibaba's Qwen 3.6 35B; one Qwen variant was post-trained by Perplexity. The Mac app handles installation without a terminal.
The pitch is twofold: privacy for sensitive work, and lower cost, since the cloud model is the expensive one and the local model is free to run. Jon Staff, who oversees Perplexity's Mac products, ties the feature to Perplexity's Computer line that began in February.
Hybrid Compute is not a privacy guarantee. The classifier is Perplexity's, with no independent benchmark of accuracy or leakage visible. For a lawyer, doctor, or analyst on a recent Mac and already using Perplexity, the split is real. For anyone who needs a hard guarantee, it is not.