Perplexity launches Hybrid Compute, which splits a task between a frontier, cloud model and a local LLM to handle sensitive info, for all users of its Mac app
Context & Ripple Effects
Perplexity introduced Computer in February as a general-purpose digital worker that could route work across 19 AI models, then outlined a Mac-focused Personal Computer in March. Its June local-and-cloud task-routing feature established the architecture Hybrid Compute is making available across the Mac app.
The rollout turns that earlier capability into a broader product setting for Mac users: a frontier cloud model can handle part of a task while a local LLM handles steps involving sensitive information. Perplexity has also said it will open-source the PII classifier used to decide that routing.
First-order effects
- Perplexity Mac app users gain access to Computer workflows that can keep sensitive task steps on-device while retaining a frontier cloud model for other steps.
- Perplexity moves its June local-and-cloud routing design from a feature announcement to a deployment for all Mac app users, making privacy-aware model selection part of the product experience.
Second-order effects
- Perplexity can shift some agent-token processing from cloud models to Macs, aligning its stated goal of preserving private data locally while improving token efficiency.
- AI agent apps built around cloud execution face a clearer product comparison on Mac: whether they can distinguish sensitive local steps from tasks suited to remote models.
Third-order effects
- If hybrid routing becomes a standard agent capability, model choice will increasingly be determined task by task by data sensitivity and available local hardware, rather than by a single cloud model.
- Open-sourcing Perplexity's PII classifier may make the privacy-routing layer itself a point of competition and reuse across hybrid AI systems.
The trend: AI agents are moving toward hybrid inference stacks that pair frontier-cloud capability with local processing for privacy-sensitive work.