NVIDIA and Perplexity on Sept. 14, 2026, made Portable Computer available in the Perplexity app for Windows on compatible GeForce RTX and RTX PRO machines. Portable Computer is an AI agent, AI that can carry out tasks on its own, and it does its work on the PC itself. It had already shipped on NVIDIA DGX Spark systems and on RTX PCs running Linux. For Pro and Max subscribers whose graphics card has enough memory, multistep agent work can now stay on their own PC instead of using up Perplexity Computer credits in the cloud for every job.
Portable Computer is Perplexity's on-device version of Perplexity Computer. It plans and runs multistep tasks with AI models on the PC, keeps sensitive files on the machine, and does not use up Perplexity Computer credits for work finished there. When a task needs web research or stronger reasoning from models in the cloud, the companies say the agent asks permission before sending anything off the device.
NVIDIA says setup uses a model that runs on the PC, such as Qwen 3.8 27B, given extra training to work with Perplexity Computer and tuned for RTX graphics cards. NVIDIA lists connectors for Microsoft Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub. Perplexity says the Windows release is available to Pro and Max subscribers, that the software coordinating the work can schedule recurring jobs, and that desktop apps can connect through local Model Context Protocol servers. NVIDIA says DGX Station support is expected soon.
According to a news report, graphics cards that meet the requirement include the GeForce RTX 3090, 3090 Ti and 4090 with 24GB and the GeForce RTX 5090 with 32GB, plus several RTX PRO Blackwell cards. Perplexity says the model, the agent software around it, and the tools that coordinate and schedule its work all run entirely on the Windows device.
It is not yet clear how many everyday RTX desktops meet the 24GB requirement without an upgrade, when DGX Station support and the RTX Spark Windows PCs due in October will widen the pool of machines that can run it, or how often users will hand tasks to cloud models in practice.