Skip to main content
Back

Perplexity Splits AI Tasks Between Your Laptop and the Cloud Automatically

Perplexity Computer will route each task between a local model on your machine and frontier cloud models automatically, keeping private data on-device.

  • Perplexity announced hybrid agentic inference for Perplexity Computer, splitting tasks between local and cloud models automatically.
  • A compact on-device model classifies sensitive data and keeps it local, routing heavier reasoning to frontier cloud models.
  • Rolling out in July on the Windows app, demoed on Intel Core Ultra Series 3 silicon.
  • Model-agnostic harness will also support NVIDIA RTX Spark systems arriving in the fall.
  • Unveiled at Computex 2026 with Intel CEO Lip-Bu Tan onstage.
  • Targets enterprise privacy needs in finance, health, and legal workloads while cutting Perplexity's own server costs.

Perplexity is bringing what it calls hybrid agentic inference to its Personal Computer agent, a system that automatically decides which parts of a task should run on a local model on your machine and which parts should be routed to frontier models in the cloud. The company unveiled what it calls the first hybrid local-server inference orchestrator at Computex 2026, demonstrating software that autonomously decides, in real time and mid-task, which AI workloads stay on a user's device and which get routed to frontier models in the cloud.

CEO Aravind Srinivas demonstrated the system onstage alongside Intel CEO Lip-Bu Tan during Intel's keynote, using Perplexity's Personal Computer agent to process confidential deal materials. The feature is coming to Perplexity Computer in July, demoed on Intel Core Ultra Series 3 processors and currently exclusive to the Windows PC app.

A router that decides where your tokens go

The orchestrator works on a simple division of labor. A compact model runs locally on your device to determine when sensitive data should be kept local, while work that needs a frontier model's full capability runs on the server. Rather than requiring users to manually choose between local and cloud processing, the system makes those decisions automatically for each request.

The pitch is that most real workloads are mixed. A request to summarize your financial records, for example, contains sensitive context (the records themselves) and a generic reasoning step (the summarization). Splitting those steps means the heavy reasoning still benefits from a frontier model while the private data never leaves the laptop.

Why this is more than another local-vs-cloud toggle

Perplexity frames this as an extension of what it already does with agent orchestration. The same harness that picks between hundreds of agents across more than twenty frontier models is now being asked to choose the hardware too: which model, where it runs, and why. The compact on-device model effectively acts as a traffic cop, classifying which information is sensitive enough to stay local before any cloud call is made.

The real shift here is treating compute itself as something to orchestrate. Most hybrid AI products today ask you to pick local or cloud up front, or run two parallel stacks. Perplexity is collapsing that decision into the same task-level routing it uses for model selection, which is what makes splitting a single task across both tiers possible.

The hardware story

Hybrid inference only works if the local chip can actually carry weight. Intel used its keynote to showcase Xeon 6+ processors with 288 efficiency cores built on 18A technology for the data center, and positioned its Core Ultra Series 3 as the client silicon that makes hybrid inference possible on the PC. Perplexity says the harness is model-agnostic and will also run on NVIDIA's RTX Spark systems, with RTX Spark systems beginning to arrive in the fall.

The business logic

There is a clear cost angle for Perplexity. CEO Aravind Srinivas framed the move around cost efficiency: Perplexity's revenue grew fivefold to $500 million while headcount rose just 34%, and offloading inference to user hardware keeps that ratio working. Every token answered by a user's own silicon is a token Perplexity does not have to pay a GPU bill for.

It also lines up with enterprise demand. IDC forecasts a tenfold increase in agent usage and a thousandfold growth in inference demands by 2027, and security and governance rank as the top evaluation factor for enterprise agentic platforms, according to a CrewAI survey. Hybrid inference speaks directly to that priority. An investment bank under data handling agreements, a hospital with patient records, or a law firm with privileged documents can keep the sensitive parsing local while still tapping a frontier model for the analytical work.

What it is good at, and what it is not

  • Good fit: mixed tasks involving private files, financial data, or health information where part of the work is generic reasoning that benefits from a bigger model.
  • Good fit: users with capable recent silicon (Intel Core Ultra Series 3, upcoming RTX Spark) who want to reduce cloud round-trips.
  • Weak fit today: Mac users, since the rollout starts on Windows only.
  • Weak fit: older laptops without a usable NPU or modern GPU, where the local model will not have enough headroom to keep meaningful work off the cloud.
  • Open question: how well the small classifier model decides what counts as sensitive. A wrong call sends private data to the server or wastes a frontier call on something trivial.

The broader signal

Industry analysts expect future AI assistants to operate across multiple layers, using local processing whenever possible and cloud resources only when necessary. If hybrid orchestration becomes the default pattern, two assumptions across the field need updating. First, that frontier-quality answers require a full round trip to a hyperscaler. Second, that the centralized GPU buildout has to scale linearly with agent usage. Pushing routine and sensitive work onto devices people already own reshapes both the privacy story and the capex story at the same time.