Perplexity has unveiled "hybrid compute," which keeps sensitive information on Mac devices while splitting other work with cloud AI. Engadget reported on Monday that hybrid compute has been integrated into the Perplexity Mac app and supports data protection and lower inference costs.

Hybrid compute splits a single task between a cloud-based model and a local large language model. Models such as Opus 5 and GPT-5.6 Sol handle the cloud side, while a model running directly on a user's computer processes the local side.

Perplexity cited as one of the main use cases a lawyer drafting a brief comparing an active case with existing precedents without sending client information outside.

John Staff, Perplexity's head of Mac products, said the app automatically checks whether content is sensitive when a file is uploaded or information is sent, and then asks again whether to send it to the cloud. Perplexity also added a feature that automatically suggests which files and information users should keep on their computers alongside hybrid compute.

Users can review which files to keep local before splitting a task through hybrid compute. They can also choose the model to use at this stage. Local models supported include Gemma E4B and 2 models in the Qwon 35B 3.6 family. One of them was further trained by Perplexity. The app installs local models without requiring users to open the Mac terminal.

Once a task starts, the screen shows local CPU, GPU and memory usage. The sidebar also shows the number of tokens used by the task. Tokens generated by a local model on a personal device are not billed. After results are produced, users can enter follow-up instructions, and they can also add tasks to the queue from an iPhone.

Perplexity said that, in terms of performance alone, a fully cloud-based approach is generally better for producing outputs. It said some users may place more importance on data protection and cost than peak performance, so it allowed users to choose based on the nature of the task.

Hybrid compute currently is supported only on Apple Silicon Macs running macOS 15. Perplexity recommends at least 32GB of unified memory. It is available to Pro and Max subscribers and enterprise customers.