Locally intercept sensitive data before transmitting to the cloud; someone has finally implemented this idea. See the real results in 2026.

CryptoWorld News: Perplexity has announced the introduction of a hybrid edge-cloud inference architecture for its desktop intelligent agent environment, Perplexity Computer. The system can automatically identify task types and privacy sensitivities, routing computation to local devices or cloud-based large models. It is expected to launch first on Intel and NVIDIA RTX Spark hardware in July 2026. With hybrid inference, tasks such as privacy scanning and personal identity information filtering are handled by local models, ensuring sensitive data does not leave the device, while complex reasoning and search tasks are offloaded to cutting-edge cloud models. In the official demo video, the local sub-agent actively intercepts after detecting sensitive files and only sends desensitized instructions to the cloud. The new architecture expands decision-making from choosing a model to determining where the computation takes place. Perplexity plans to integrate local and cloud computing power, running low-cost tasks locally, only to offload expensive tasks to the cloud.