The appealing part of a capable AI agent is also the alarming part: it needs to see the work around it. Emails, spreadsheets, repositories and folders become the raw material. Perplexity’s Portable Computer now brings that proposition to compatible Windows RTX PCs, with a local model doing the first pass instead of sending every task straight to a cloud service.
That is a meaningful change for people whose work is too sensitive or too sprawling for a browser chat window. NVIDIA says the Windows version can use local models to work across files and recurring tasks, while asking before a job needs cloud-level reasoning. Locally completed work does not consume Perplexity Computer credits, according to the company.
The catch is refreshingly concrete: this is not an “AI PC” feature for every recent Windows laptop. NVIDIA lists a GeForce RTX or RTX PRO GPU with 24GB or more of VRAM. That puts the opening act in the world of high-end desktop cards, mobile workstations and specialist setups.
What Windows support actually changes
Portable Computer debuted on Linux and NVIDIA’s DGX Spark systems in August. VentureBeat and Tom’s Guide both described that initial release as local-first: tasks begin on the device, and the user is meant to approve any step that must move to a frontier model in the cloud. The Windows release removes the operating-system detour for the much larger group already working from RTX-equipped PCs.
NVIDIA says the app can connect with Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub. Those integrations are why the local model matters. A useful agent needs more context than a single pasted prompt, but a connector is also a permission boundary—not proof that every workflow is private by default.
| Question | What the release says | The practical reading |
|---|---|---|
| Where does work start? | On the PC with a local model | Sensitive files can stay local for work the model can complete there. |
| When is the cloud involved? | When more advanced reasoning is needed, after permission | Check the prompt carefully; a local-first workflow is not a cloud-free guarantee. |
| Which hardware qualifies? | NVIDIA GeForce RTX or RTX PRO with 24GB+ VRAM | Typical thin-and-light laptops are out. |
| Which models are mentioned? | A Perplexity-tuned local model such as Qwen 3.8 27B | The experience depends on the local model’s limits as much as the agent UI. |
| Is it a standalone offline office suite? | No | Connected services and optional cloud escalation remain part of the proposition. |
The privacy win is real, but conditional
The most sensible use case is not asking an agent to write a cheery email. It is asking it to sift a folder of contracts, trace a number across reports, triage a repository, or make sense of a messy local export without first uploading the full archive to a generic chatbot.
That is a better starting position for private work. It does not erase ordinary security hygiene. Users still need to decide which accounts and connectors to authorize, what data an agent may act on, and whether a particular task is worth escalating. “Ask before cloud” is a good control; it is not the same thing as an audit of every downstream action.
The local-model requirement also brings a quieter trade-off: smaller local models can be fast and private, yet less capable on the hard reasoning steps that tempt people to click the cloud option. The right expectation is a capable assistant for bounded, inspectable workflows—not an unattended colleague with access to the company drive.
Why the 24GB floor matters more than the word “Windows”
The release looks broad because Windows is broad. The hardware support is not. NVIDIA’s 24GB threshold is a useful reality check on the local-agent conversation: real multistep work with a 27-billion-parameter class model still needs substantial graphics memory.
That makes Portable Computer more relevant today to developers, analysts, creative professionals and teams that already own capable RTX hardware than to anyone shopping for an everyday Copilot-style laptop. Tom’s Guide made the same point before the Windows release: the hardware constraint keeps the first version closer to enthusiasts and organizations than mainstream consumers.
There is still value in that. Plenty of workstations already have the GPU headroom, and the industry has been steadily turning that headroom into software rather than just rendering or gaming performance. Our look at NVIDIA’s Nemotron enterprise push made a related point: the interesting competition is shifting from the model alone to where an agent runs, what it can touch, and who controls the bill.
Who should care now
- Worth investigating: teams with 24GB+ RTX hardware, sensitive document workflows, and a clear policy for cloud escalation.
- Worth waiting on: most laptop owners, people without a use case beyond casual chat, and anyone who needs a fully offline tool with no external connectors.
- Worth testing cautiously: developers using GitHub or operations teams using Drive, Slack and Outlook. Start with a copy of non-critical data and inspect proposed actions.
Bottom line
Perplexity Portable Computer on Windows is a serious sign that local AI agents are leaving the demo-room phase. Its most persuasive idea is not free local compute; it is giving the user a chance to keep the boring, sensitive middle of a task on their own machine.
For now, though, the 24GB VRAM requirement is the headline beneath the headline. This is a promising workstation tool, not a reason to believe every Windows PC has suddenly become a private AI powerhouse. If you already own the right RTX hardware, it is worth a measured trial. If you do not, wait for the hardware floor to come down before treating local agents as an everyday feature.