For makers and artists, the era of choosing between privacy and performance is ending. Perplexity AI has unveiled a new system that automatically decides which AI tasks run on your machine and which require the power of the cloud. This hybrid approach means your creative workflows can leverage massive frontier models without ever sending sensitive drafts, financial records, or proprietary code to external servers. The technology arrives in Perplexity Computer later this year.
The logic behind hybrid agentic inference
Developing this system required resolving a fundamental conflict in modern AI architecture. High accuracy demands massive models, which are computationally expensive. Privacy dictates that certain data must never leave your hardware. Meanwhile, cost efficiency requires avoiding the use of supercomputers for simple queries.
This routing layer is what Perplexity terms hybrid agentic inference. A compact model operates locally on your device, acting as a gatekeeper. It evaluates every incoming request to determine if it involves sensitive information, requires heavy processing, or can be handled entirely offline. Based on this assessment, the system either processes the work locally or dispatches it to a frontier model in the cloud.
Crucially, the local model decides when sensitive data should remain on-device. The architecture explicitly asks for user permission before transmitting sensitive tasks to the cloud. This design directly addresses enterprise concerns regarding data governance—ensuring users know exactly where their data travels and who controls that decision.
The system is designed to keep financial records, health information, and personal files strictly local. Conversely, tasks requiring the full capability of a frontier model execute on the server. Since most real-world projects are a mixture of both, the system splits the workload and coordinates the parts seamlessly.
Integrating into Perplexity Computer
Perplexity Computer is the company’s existing cloud-based multi-model agentic product, launched in February 2026. It originally operated entirely in the cloud, accessible via the Perplexity Max subscription tier ($200/month).
Personal Computer is a distinct, related product that brings those capabilities to the local device. It offers access to local files, native Mac applications, the web, and Perplexity’s secure servers. Personal Computer launched on Mac in April 2026, with Windows support currently on a waitlist.
The new hybrid local-server inference orchestrator represents the next evolution for Personal Computer. Previously, the division of labour was relatively fixed: local file access occurred on-device, while heavy computation ran on Perplexity’s servers. The orchestrator changes this dynamic. The system now reasons about where each piece of a task should execute—determining not just which model to use, but the physical location that should process it.
Perplexity Computer coordinates up to 20 AI models within a single workflow. It creates a team of agents to orchestrate across models, tools, and files. This hybrid orchestrator extends that coordination to include compute location itself, ensuring the right hardware handles the right load.
Key takeaways
- Perplexity AI announced the first hybrid local-server inference orchestrator at Computex 2026, enabling automatic routing of AI tasks between on-device and cloud models.
- A compact local model acts as the router — classifying each subtask by data sensitivity and compute requirements before dispatching it.
- Sensitive data (financial records, health files) stays on-device; compute-heavy tasks go to frontier cloud models — no manual configuration required.
- The orchestration framework is model-agnostic and chip-agnostic, confirmed to run on Intel Core Ultra Series 3 and NVIDIA RTX Spark hardware.
- The feature arrives in Perplexity Computer in July 2026, initially on Windows; Personal Computer is already available on Mac with a Windows waitlist open.
Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.
