Perplexity has launched the Hybrid Compute feature for its agent platform, Computer, in an effort to combine the capabilities of advanced cloud models with the privacy of local processing. The feature allows the agent to start a task in the cloud and then delegate portions involving private files or data stored on the device to a model running locally on a Mac equipped with an Apple silicon chip, without restarting the task or losing its context.
The feature is available starting September 1, 2026, through the Perplexity desktop app for enterprise customers who choose to enable it, as well as for Pro and Max subscribers, on any Mac with an Apple silicon processor running macOS 15 or later. The company did not specify in the material a particular date for Windows or Linux support, saying only that support would come later.
How Does the Platform Distribute Task Components?
The architecture operates as a coordination layer, or task router. An advanced cloud model breaks the request into subtasks and then decides where each should be executed. Web research, long-term planning, and intensive reasoning tasks can remain in the cloud, while operations that touch private files, local data, or device actions are sent to a sub-agent running on the Mac.
Perplexity relies on what it calls Privacy Gate, a classifier trained by the company that runs on the device to scan for personally identifiable information, such as names, addresses, account numbers, and secrets, before any content is transferred to the cloud. When sensitive data is detected, the user chooses whether the relevant portion will be executed locally or shared with the cloud coordinator. The company says that tokens generated by local execution do not leave the device.
Local Models and a Different Cost Structure
At launch, users can choose among three local models: Gemma E4B from Google, Qwen3.6 35B-A3B from Alibaba, and a version of Qwen3.6 35B retrained by Perplexity, which is the model the company recommends. Jon Staff, who leads Perplexity’s macOS and iOS engineering teams, explained that running a Qwen model locally does not send tokens to a cloud provider hosted in another country, and noted that the company’s models are hosted in the United States.
Tokens generated by local execution do not count against usage credits, according to Staff, while credits are used for cloud coordination and delegation. However, Perplexity recommends at least 32GB of unified memory to run the more powerful class of local models, and Staff acknowledged that the smaller model performs significantly worse than the larger Qwen models.
What Changes Practically for Enterprises?
Perplexity presented use cases aimed at work involving data that is difficult to send in full to a cloud agent. In a legal example, the system processed confidential case files locally while searching public legal precedents on the web and, according to the company, sending only anonymized legal questions. In another example, an agent processed a financial model based on confidential management forecasts, compared it with public data, and then created a fifth version of an investment committee presentation in the background in about 40 minutes.
Users can also start a task from an iPhone and give Computer permission to access a remote Mac, allowing the local device to process customer interviews and revenue data while the cloud agent searches for public competitor pricing. For enterprises, Perplexity allows administrators to apply a unified sensitivity policy across the organization and audit a complete record of what leaves each device—features aimed particularly at compliance teams in the legal, financial, and healthcare sectors.
Why Does This Launch Matter?
The product attempts to address a practical trade-off: work requiring the greatest accuracy often depends on data that organizations do not want to send to external servers, while local models alone may be less capable of research, planning, and reasoning. Perplexity therefore does not present hybrid processing as a complete replacement for the cloud or the local device, but rather as a layer that decides where each part of a task should be executed.
However, this protection is not an absolute guarantee. The privacy gate itself is a machine-learning classifier and may fail to detect some sensitive data. The company says users can review what the gate identified before sending it, while organizations receive device-level audit logs, but the mechanism’s practical effectiveness will depend on the classifier’s accuracy and on the review policies established by each organization. Details about the use of data from non-enterprise accounts for model training also remained in need of further clarification; the company said it does not use such data for global post-training and promised to provide details for non-enterprise accounts.