Perplexity equips Computer with hybrid inference between the device and the cloud

Perplexity Computer adds hybrid agentic inference to split tasks between local chips and cloud models, boosting privacy and energy efficiency by July.

Perplexity is expanding its Computer product with hybrid agentic inference, an orchestrator that automatically distributes tasks between a model running locally on the machine and state-of-the-art models hosted in the cloud. A compact model runs on the device and determines which sensitive data, such as financial records, health information, or personal files, should remain there, while processing requiring the full power of a large model is offloaded to the server.

For most tasks combining both, the system splits them and coordinates their parts, without requiring the user to decide between local and remote upfront. The stated goal is to optimize the value obtained per watt consumed, by arbitrating between accuracy, privacy, and energy cost according to the needs of each task. The more powerful the local chip is, the more computation the orchestrator can keep on the device and reserve the cloud for work that truly requires it. Presented with Intel, the infrastructure aims to be hardware-agnostic and runs on other chips, including NVIDIA's RTX Spark. Perplexity also sees it as a way to reduce reliance on centralized data centers and keep certain data within its original jurisdiction. The local inference variant, named Personal Computer, is expected in July.