Perplexity's Portable Computer, the local version of its Computer agent that plans and carries out multistep tasks, is now available in the Perplexity app for Windows on compatible NVIDIA GeForce RTX PCs and RTX PRO Workstations. The requirement is a card with 24GB of VRAM or more, and NVIDIA says DGX Station support is expected to arrive soon. The release extends an agent that already ran on DGX Spark systems and RTX PCs under Linux.
The pitch is privacy plus cost. Work the agent completes locally stays on the machine, and it does not draw down Perplexity Computer credits, according to the announcement. When a task needs heavier reasoning, Portable Computer can recognize that and ask the user before sending anything off the device to cloud models.
Read more: Gemini App Lands on Windows With Alt + Space Shortcut
Setup is meant to be simple. The app ships with a local model, such as Qwen 3.8 27B, post-trained to work with Perplexity Computer and tuned for RTX GPUs, so users do not have to pick models or assemble the software stack local AI usually demands. Connectors for Microsoft Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub pull the agent into files and apps people already use.
NVIDIA's examples sketch out what that looks like in practice. An engineering team could have the agent sort open GitHub pull requests by status, flag stale documentation and propose updates for review. A finance user could point it at two years of brokerage summaries, consolidated 1099s and tax returns to trace recurring holdings behind avoidable fees and tax drag, with each figure cited to a file and page and no document leaving the machine. A startup could ask why activation flattened and get a local read of the funnel export, with the top findings posted to Slack.
The Windows launch follows NVIDIA's broader local AI push at IFA 2026, where the company and Microsoft lined up partners behind compact RTX Spark Windows PCs due in October. At that time Portable Computer was limited to RTX GPUs with at least 24GB of VRAM on Linux, with Windows support described as coming soon. The same event brought one-click local model setup for Hermes Agent on Windows and an OpenClaw Windows app for RTX GPUs with 24GB of VRAM or more.
What the announcement does not include is a price for Portable Computer itself, or any statement about whether the Windows release carries different hardware requirements than the Linux version beyond the 24GB VRAM floor. NVIDIA also lists a batch of unrelated local model updates in the same post, including Z.ai's GLM 5.3 Flash, DeepSeek-v4.1 Flash and an early Qwen4 preview that runs on a single DGX Spark.
Perplexity is positioning the agent as a way to work with sensitive material without handing it to a chatbot, a claim that rests on the local processing and the permission prompt before any off-device transfer. How widely that applies is limited by the 24GB VRAM floor NVIDIA lists for the release.
