Perplexity launches local AI agent Portable Computer with Nvidia

Perplexity has launched Portable Computer, a version of its AI agent platform that runs on a user’s own Nvidia equipped hardware rather than primarily in the cloud. Michael Nuñez reports for VentureBeat that the product was developed with Nvidia and is initially available for Linux systems with supported RTX graphics cards and Nvidia’s DGX Spark desktop system.

The software is designed to handle multi step work such as reviewing documents, analysing spreadsheets, creating reports and using connected business services. Perplexity says the model, user files, tools and AI workflow can remain on the device. That means locally completed tasks do not consume cloud credits or incur per token charges.

Portable Computer starts work locally by default. Users can choose to send a specific step to a more capable cloud model if a local model cannot complete it. Before such a request, Perplexity says it checks the context for personally identifiable information and shows users what data would leave the device. The remote model only returns guidance and cannot directly access local files or tools.

Built for local work

At launch, users can install Qwen 3.8 27B or Perplexity’s post trained PPLX 27B model. The company requires an Nvidia RTX GPU with at least 24GB of video memory, which limits support to higher end systems such as the GeForce RTX 3090 and newer hardware. Windows support is planned for September.

Perplexity argues that smaller local models need a simpler agent framework than frontier cloud models. It has reduced the number of tools held in a model’s context, loaded capabilities only when needed, and added operating system level sandboxing for tool use. The company reports stronger results than two open source agent frameworks in its own tests, though those figures have not been independently verified.

The launch positions local AI as an option for organisations handling sensitive material or seeking more predictable AI costs. However, local models still lag behind leading cloud systems on difficult reasoning tasks, and the hardware requirements remain high.