Perplexity and Nvidia launch a fully local AI agent, Portable Computer

Perplexity and Nvidia released Portable Computer, a version of Perplexity’s AI agent that runs entirely on local hardware rather than calling a cloud model for every step. It works on Nvidia’s DGX Spark desktop system or on a Linux machine with an RTX GPU carrying at least 24GB of VRAM — roughly an RTX 3090 or newer — using Perplexity’s own PPLX 27B model or Alibaba’s open-weight Qwen 3.8 27B, with Nvidia’s Nemotron 3.5 Lightning described as coming later.

Because inference happens on the user’s own machine, VentureBeat reported, tasks completed locally draw no cloud billing credits; Perplexity said the interface is unchanged from its cloud product, and the system can escalate a task to a larger cloud model, such as Anthropic’s Claude Opus 5, when local hardware cannot handle it. The company framed the escalated case as cheaper too, citing a lower per-task cost for the hybrid route than for running everything on a frontier model in the cloud. The stack also packages agent orchestration, tool access and sandboxing locally, including operating-system-level isolation and a check on personal data before anything is sent to the cloud, with integrations for Google Drive, Gmail and GitHub.

The release is one of several 2026 attempts, alongside efforts from other labs and from open-weight model providers, to shift agentic AI work off metered cloud inference and onto a user’s own GPU — trading a hardware cost for a recurring one, and, for users already holding Nvidia workstation or gaming cards, removing the token meter entirely for tasks the local model can handle.