Perplexity has expanded its local AI agent, Portable Computer, to Windows machines equipped with NVIDIA GeForce RTX GPUs and RTX PRO workstation cards, according to an NVIDIA blog post published September 14. The agent previously ran only on NVIDIA DGX Spark systems and RTX equipped Linux machines.

Portable Computer is a scaled down version of Perplexity Computer, the company’s agent for planning and executing multistep tasks. Running locally on a device’s own GPU lets the software keep sensitive documents off remote servers and skips the usage credits Perplexity charges for cloud based Computer sessions, per NVIDIA’s writeup.

The Windows release requires a GeForce RTX or RTX PRO card with at least 24GB of VRAM, a threshold that excludes most consumer laptops and mid range desktop cards still on the market today. NVIDIA said support for its DGX Station hardware is coming, without giving a date.

To run without heavy setup, the agent ships with a locally hosted model, Qwen 3.8 27B, that NVIDIA said Perplexity post-trained specifically for the Computer workflow and tuned for RTX hardware. When a task exceeds what the local model can reasonably handle, Portable Computer can ask the user for permission before routing the request to Perplexity’s cloud models, according to the post.

Perplexity built connectors into the tool for Microsoft Outlook, OneDrive and Word, Google Drive, Gmail, Slack and GitHub, letting the agent reach into a user’s existing files rather than requiring documents to be re-uploaded into a chat window. NVIDIA’s post walks through three use cases: an engineering workflow that sorts open GitHub pull requests and flags outdated docs, a finance task that traces recurring account fees across years of brokerage and tax records, and a startup analysis that digs through a funnel export to locate where signups stall, then posts findings to Slack.

Those examples come from Perplexity and NVIDIA themselves, not independently verified results, and neither company published benchmark data showing how the local model’s output compares with the cloud hosted version of Computer on identical tasks.

The push toward on-device agents reflects a broader shift in how AI companies position local compute: not as a downgrade from the cloud, but as the default choice for anything involving financial statements, health records, or unreleased code a company does not want leaving its own hardware. Anthropic, OpenAI and Google have each shipped lighter local or on-device assistant variants over the past year, though few set a VRAM floor as high as the one Perplexity is setting here.

For operators evaluating agentic tools that touch regulated or proprietary data, the practical filter is now hardware, not policy. A 24GB VRAM minimum puts Portable Computer out of reach for most laptops sold today, so teams weighing it against cloud only agents should budget for a GPU refresh alongside any software rollout.

Reported by NVIDIA on its company blog, September 14, 2026, describing Perplexity’s own product release.