Perplexity has released a hybrid compute feature for Mac that splits agentic tasks between cloud frontier models and local models, using an on-device privacy gate to handle sensitive data. This allows users to leverage powerful cloud reasoning while keeping private files like deal documents or client records strictly on their hardware.
- The system starts tasks in the cloud but hands off steps touching sensitive data to a local model without losing context.
- An open-sourced 0.6B classifier, PII-Tracer, inspects data before it leaves the device, applying outcomes like masking or refusal.
- PII-Tracer leads 12 detectors on character F1 (0.629) and consistency for recurring identifiers (79.4%).
- Launch includes local models Gemma 4 E4B, Qwen3.6 35B-A3B, and a Perplexity post-trained model.
- Available for Pro, Max, and Enterprise subscribers on Apple silicon Macs with macOS 15+ and 24GB+ unified memory.
This architecture addresses the structural problem of agentic assistants by enabling secure processing of privileged information that users cannot send to cloud endpoints.