Perplexity announced the launch of a new feature called Hybrid Compute, which allows users to split a task between cloud-based cutting-edge models and local large language models - with sensitive information handed over to the local model to ensure that related data always remains securely on the user's own device.

image.png

Upload and Detect Sensitive Content, Install Local Model with One Click

Jon Staff, Product Manager of Perplexity Mac, introduced this feature, which is directly integrated into the Mac application: "Every time you try to upload a file or send a message, the system automatically checks whether it contains sensitive content and confirms whether you are willing to share this data with the cloud."

To this end, Perplexity has also trained a new privacy classifier, which can automatically suggest files and information that should be kept on the computer. Before assigning the task to different models, users can first view the files prepared for isolation and kept locally by the system, confirm there are no important omissions, and decide which models will complete the task.

Currently available local models include Gemma E4B, as well as two versions of the Qwen 3.6 model, both of which have 35 billion parameters, with one of them being post-trained by Perplexity. The company said more local models will be added in the future. No matter which one is selected, the installation process does not require opening the Mac terminal, as it is automatically completed by the Perplexity application. During operation, users can see visual information about the usage of the local CPU, GPU, and memory; the sidebar shows the number of tokens consumed by the task; tokens generated by the local model do not incur costs, and after the result is produced, subsequent instructions can still be input, and tasks can also be queued through the iPhone.

Cloud Is Stronger but More Expensive, It's a "Tunable Scale"

When asked if they had benchmarked the results of hybrid computing against pure cloud solutions, Staff admitted, "In short, if we look only at the quality of the original output, the output from using the cutting-edge model alone is almost always better. It is more expensive, but its capabilities are stronger." However, he added that not all users need the most advanced and powerful model to complete their work, and in such cases, data privacy and cost may be more important than model capability.

"I think this should be a tunable scale... We will try to recommend the most suitable solution based on the context we have, but ultimately, the user should decide which solution best fits their needs," Staff said.

Currently, Hybrid Compute is only available for Mac computers equipped with Apple Silicon chips running macOS 15, and Perplexity recommends that devices have at least 32GB of unified memory. This feature is available to Pro and Max subscription users and enterprise customers.