Perplexity on Tuesday released Hybrid Compute, a feature that allows its AI agents to divide individual tasks between powerful cloud-based models and smaller models operating directly on Apple silicon inside a Mac. The launch coincides with John Ternus' first day as Apple's CEO, succeeding Tim Cook after 15 years at the helm. Perplexity is wagering that the same hardware Ternus helped develop—he previously ran Apple's hardware engineering division and steered the shift from Intel chips to Apple silicon—can now serve as infrastructure for autonomous AI agents.
The system begins operations in the cloud, but when a step involves sensitive details, Perplexity Computer can transfer that portion of the work to a model running on the Mac without restarting the entire task. A Perplexity-trained Privacy Gate operates locally on the Mac, scanning for sensitive data like names, addresses, account numbers, and secrets. When it detects something, the user decides whether that segment should remain on the device. Users can review what the system plans to keep local before work starts and verify it hasn't overlooked anything they don't want transmitted to the cloud. At launch, users can select from Gemma E4B, Qwen3.6 35B-A3B, and a version of Qwen3.6 35B that Perplexity post-trained itself, with additional models planned for the future. Users aren't billed for tokens generated by models operating locally on their Mac. The app shows local CPU, GPU, and memory use, along with the number of tokens consumed during a task.
The report notes that Perplexity deploys its most capable models in the cloud, while smaller models manage local operations, so retaining more of a task on the Mac can boost privacy and reduce costs at the expense of some capability. According to the announcement, Computer can shift a step from the cloud to a local model without restarting the task or losing context, and ultimately merge the cloud and local work into a single result. The local subagent can interact with private files and data and execute actions on the Mac, and users can start a task on an iPhone and hand off local work to their Mac without starting over.
The approach addresses two emerging pressures in AI deployment: who pays for inference and where sensitive information gets processed. By moving work onto the device, Perplexity cuts its own inference costs while giving users control over what leaves their machine. The cloud model still needs enough information about what occurred locally to continue the job, without accessing the private details that were meant to stay there. Perplexity doesn't spell out in its announcement exactly what context passes between those environments or how information produced by the local subagent gets filtered before returning to the broader workflow. For enterprise customers, Perplexity adds company-wide rules for what stays local and a record of what exits each device. The feature is available to Pro and Max subscribers as well as enterprise customers, but requires at least 32GB of unified memory, Apple silicon, and macOS 15—showing that giving an agent enough compute to handle meaningful work still demands relatively high-end hardware. The report suggests that will likely change as Macs improve at running larger models, with the bigger question for Ternus being how much of the AI work now occurring in the cloud will eventually shift onto the machines Apple sells. The orchestration layer—deciding which tools and context an agent actually needs—is becoming an increasingly difficult engineering challenge as agents gain access to more systems and data. For now, the balance between privacy, cost, and capability remains a user choice, but the infrastructure is positioning local hardware as a viable alternative to pure cloud execution.