Perplexity Portable Computer, the company’s local AI agent, is now available on Windows PCs built around an NVIDIA RTX GPU. The release brings to Windows a capability that until now existed only on Linux, delivered through the existing Perplexity app for Windows, in partnership with NVIDIA, and it arrives with a hard line drawn at 24GB of VRAM.
The pitch is privacy and cost. NVIDIA frames the shift as local models becoming capable enough to handle more agent work directly on a PC while keeping sensitive information on the device rather than sending it to a server. Work done locally does not consume Perplexity’s cloud-based Computer credits, though the agent can still hand a task to cloud models with a user’s explicit permission, according to the company.
What Perplexity Portable Computer does locally
Portable Computer is the local version of Perplexity Computer, the company’s agent for multistep tasks that plans a job, runs it through connectors and tools, and returns a finished result instead of a chat reply. That output, an edited document or a filled spreadsheet rather than a wall of text, is the mechanism the cloud version already used; what changes on Windows is where the reasoning happens. On Windows that adds scheduled recurring tasks and local MCP servers that let the agent reach desktop apps directly. NVIDIA’s rundown of the release lists connectors for Microsoft Word, Google Drive, Gmail, Slack and GitHub.
A dropdown in the app downloads a local model in one click; NVIDIA named Qwen 3.8 27B as an example. Portable Computer first launched for Linux on the DGX Spark on 25 August, and NVIDIA said on 3 September that Windows support was coming.
The 24GB VRAM gate cuts across NVIDIA’s lineup
Running Portable Computer on Windows requires a GeForce RTX or RTX PRO GPU with at least 24GB of VRAM, plus a Pro or Max Perplexity subscription. The subscription requirement applies even though the compute runs locally, since Portable Computer is a feature of the Perplexity app rather than a standalone download. That VRAM threshold behaves as a memory gate rather than a generation gate: it reaches back to older cards like the RTX 3090 and 3090 Ti as readily as it covers NVIDIA’s newer RTX PRO workstation cards, and it shuts out most mainstream consumer GPUs sold today regardless of how new they are.
Vendor framing versus the fine print
NVIDIA’s own announcement leans on words like “available” and “powered”, the language of a product now shipping. The restrictive part of the story, the 24GB VRAM floor and the Pro or Max subscription requirement, doesn’t get the same billing. Tom’s Hardware’s coverage is where the compatibility gate gets stated plainly, along with a quote from Perplexity CEO Aravind Srinivas, who wrote on X on 14 September that the release brings “unmetered local intelligence on every Windows PC running on Nvidia hardware and Perplexity harness.” Unmetered local compute is accurate for owners of qualifying cards; it says nothing about how few Windows RTX systems actually clear 24GB.
What to watch next
NVIDIA said DGX Station support is coming soon, which would extend Portable Computer to NVIDIA’s other local AI workstation beyond the DGX Spark it launched on. Whether that support keeps the same 24GB floor, or scales with DGX Station’s own memory configuration, hasn’t been said yet.
