Your files stay put: Perplexity’s hybrid AI keeps confidential data off the cloud

Perplexity introduced hybrid compute for its agentic platform, Computer, a system that enables AI agents to split work between cloud-based frontier models and smaller locally-running open-weight models on Apple silicon Macs. This feature ensures sensitive data remains on the device and never leaves the local machine. The company claims to be the first to enable such dynamic, context-preserving task-handing between cloud and local models without restarting or losing context.

The hybrid compute is available to enterprise customers and Pro/Max subscribers on Apple silicon Macs running macOS 15 or later.

Jon Staff, leading Perplexity's macOS and iOS engineering teams, emphasized the importance of confidentiality in tasks requiring high accuracy. By combining cloud and local models, Perplexity achieves maximum intelligence while maintaining security and privacy. The architecture functions like a dispatcher, with a cloud frontier model breaking a task into subtasks and routing them to appropriate places.

Web research, long-horizon planning, and heavy reasoning occur in the cloud, while tasks involving private files, local data, or device actions are delegated to a subagent on the Mac.

A key component of this system is the Privacy Gate, a company-trained classifier that scans for personally identifiable information (PII) before transmitting data to the cloud. When the gate detects sensitive content, users can decide whether the task runs locally or is shared. Perplexity built and trained its own PII classifier directly into the Mac app to ensure secure data handling.

Perplexity demonstrated the hybrid compute's capabilities through examples involving a lawyer updating a brief with privileged case files on a Mac, a private equity associate working on confidential financial projections, and a pottery shop founder creating a marketing analysis from an Uber. These demonstrations highlight the system's ability to maintain security and privacy while leveraging the intelligence of frontier models.

Perplexity offers three local models: Google's Gemma E4B, Alibaba's Qwen3.6 35B-A3B, and a post-trained version of Qwen3.6 35B from Perplexity, which users can choose from. The company reassured that these locally running models are open weight, negating geopolitical concerns, and their hosted servers are based in the United States.

Additionally, they utilize macOS's built-in sandboxing framework, Seatbelt, to ensure safe local execution and provide audit trails for compliance purposes.