Perplexity Computer adds hybrid inference to split tasks between local and cloud models

Perplexity announced that its Computer agentic system will gain the ability to automatically split tasks between on-device and cloud-based models in July 2026.

The hybrid inference feature allows Perplexity Computer to route different parts of a single task to appropriate models: sensitive data processing stays on local models running on the user's device, while complex operations requiring frontier model capabilities run on cloud servers.

How hybrid agentic inference works

According to Perplexity, a compact model running locally determines which data should remain on-device. The system then coordinates between local and cloud execution for different parts of the same task.

Perplexity positions this as solving a key problem with current AI systems: "Unlike tools that ask you to pick local or cloud up front, this happens on its own, task by task."

The company cites use cases involving financial records, health information, and personal files as scenarios where users need both privacy and powerful AI capabilities.

Technical details

Perplexity claims the hybrid approach "maximizes token efficiency" by routing work to appropriate compute resources. The announcement did not specify which local models will run on-device, what minimum hardware requirements exist, or which frontier models will handle cloud-side processing.

No pricing changes were announced for the feature, which will be available to Perplexity Computer subscribers.

Availability

The hybrid agentic inference capability will launch in July 2026 for Perplexity Computer on Mac. The company has not announced plans for other platforms.

Perplexity Computer, the company's agentic AI system, launched earlier in 2026 as a Mac-native platform for subscribers of Perplexity's paid service.

What this means

This represents a practical approach to the privacy-versus-capability tradeoff in AI applications. By automating the decision of what runs where, Perplexity aims to eliminate a manual step that many users lack the technical knowledge to handle. The success of this approach depends on the accuracy of the local model's routing decisions and the actual performance difference users experience compared to cloud-only execution. If the local routing logic proves reliable, this could become a standard architecture for privacy-sensitive AI applications.