Agentic AI assistants face a fundamental privacy challenge: the very context that makes them useful—deal documents, privileged files, client records—is often exactly what users cannot upload to cloud servers. Perplexity has now introduced a solution for Mac users. Its new hybrid compute approach splits a single task from the Perplexity Computer between frontier models running in the cloud and a compact local model on the user's Mac. An on-device privacy gate determines what data may cross that boundary. Perplexity has also open-sourced the classifier powering this gate.
Deployment and Requirements
Hybrid compute is currently available to all Pro, Max, and Enterprise subscribers using Apple silicon Macs running macOS 15 or later, with at least 24GB of unified memory (32GB recommended). Installation is streamlined: the local model downloads with a single click from the Mac app—no Ollama, no separate runtime, and no API key required. Local processing does not consume cloud credits.
How Hybrid Compute Works
The orchestration flow is intentional. Every task starts in the cloud, where frontier models handle web search, planning, and complex, multi-step reasoning. When a step involves private or sensitive files, the system automatically delegates that step to the local on-device model, keeping sensitive data off external servers. The privacy gate acts as a real-time guardian, ensuring that only non-sensitive data is transmitted.
Open-Sourced Privacy Classifier
Perplexity has open-sourced the classifier that powers this privacy gate, allowing developers and security researchers to inspect, audit, and contribute to the decision-making logic. This move underscores a growing emphasis on transparency in AI systems handling user data.
Implications and Future Outlook
This hybrid architecture represents a meaningful shift in how AI assistants can balance capability with data protection. By enabling cloud-grade intelligence to work seamlessly with local execution, Perplexity addresses a core barrier to enterprise adoption—privacy without sacrificing performance. As we move into 2026, we can expect more AI platforms to adopt similar hybrid models, making on-device compute an integral part of the user experience.
For more details on the feature and to see the open-source release, visit the official Perplexity blog.
via MarkTechPost