Perplexity has launched a hybrid processing feature for its agent platform Computer. The system allows a single AI agent to divide tasks between advanced models in the cloud and smaller open-weight models running locally on Apple silicon Macs. Sensitive data does not leave the device.
The company says it is the first system to allow a task to begin in the cloud and then transfer confidential portions within the same task to a model on the user’s hardware without restarting the work or losing context. The feature is available on Apple silicon Macs running macOS 15 or later, on an opt-in basis for enterprise customers and Pro and Max subscribers.
The architecture works by having a model in the cloud divide the task into subtasks and route them. While web research, long-term planning and intensive reasoning are carried out in the cloud, private files, local data and on-device operations are transferred to the sub-agent on the Mac. Perplexity’s on-device Privacy Gate classifier scans for personal information such as names, addresses, account numbers and secrets. If sensitive content is detected, the user decides whether the operation should be performed locally or shared.
There is no charge for locally generated tokens; credits are used only for cloud coordination and routing. Demonstrations showed a lawyer researching publicly available case law while working with privileged case files, as well as a private equity analyst updating the model with confidential financial projections and preparing the fifth version of an investment committee presentation.
Why it matters
This structure introduces a task-based sharing model for AI agents, rather than forcing a choice between cloud capacity and on-device privacy. For users working with sensitive content such as legal files and financial projections, the key point is that the decision over which portion of the data will leave the device is left to the user. Although Privacy Gate provides an additional layer of control by separating personal information, it is not clearly stated how responsibility will be allocated when the classification is incorrect. Restricting use to specific operating system, hardware and subscription requirements also limits the feature’s accessibility. The absence of charges for local operations means that cloud coordination and routing usage will be the primary factors in calculating costs.