Perplexity Portable Computer: Local AI Lands on Windows PCs

The Perplexity Portable Computer arrived on Windows PCs on September 14, 2026, and it moves the whole agent onto your own hardware. Model, planner and scheduler all run on the machine in front of you. However, there is a price of entry, and it is measured in video memory rather than dollars.

Perplexity Computer is the company’s agent for multistep work: reading files, pulling numbers out of them, and producing a document at the end. The Portable Computer is that same agent, packaged to run entirely on hardware you own.

Until now it lived on Linux boxes and on NVIDIA’s DGX Spark. As of September 14, 2026, it installs on Windows through the Perplexity app in the Microsoft Store. Consequently, a far larger group of people can try local agent work without buying a dedicated machine.

According to NVIDIA’s announcement, the model, the agent harness, the orchestrator and the scheduler all execute on the device. Nothing about the loop depends on a data centre.

The practical draw is automation that does not leak. The agent plans a task, opens your local files, and works through spreadsheets, slide decks, PDFs, code and images. Meanwhile, the documents never leave the disk they sit on.

It also reaches into the apps people actually live in. Perplexity lists Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub among the connections. In addition, you can schedule a workflow to repeat, so a weekly analysis simply runs itself.

Perplexity’s VP of engineering framed the port bluntly: “We’ve basically brought the exact same UI to a fully local app.” The company says the local build incorporates the entirety of the agent harness and inference.

Here is where most PCs fall out. The requirement is an NVIDIA GeForce RTX or RTX PRO card carrying at least 24GB of VRAM. In practice that means an RTX 3090 or something newer, plus support for the RTX PRO workstation line.

Perplexity also says NVIDIA DGX Station support is coming. The original launch target, the compact NVIDIA DGX Spark, was built for exactly this workload. Because of that heritage, the Windows build inherits a fairly narrow hardware list.

The table below collects what the company has specified so far.

The app itself is not a free tier product. Portable Computer goes to Pro and Max subscribers, on both individual and enterprise plans. Instead of charging per task, though, Perplexity changes what a task costs after that.

Work the agent finishes locally does not spend Perplexity Computer credits. As a result, a long batch job on your own GPU is effectively free to repeat. That is a genuinely different economics from a cloud agent metered by the token.

Small models hit walls, and Perplexity does not pretend otherwise. The agent runs Qwen 3.8 27B, post-trained for Perplexity Computer and tuned for RTX cards. NVIDIA’s Nemotron 3.5 Lightning, a 30B model, is queued to join it.

When a job needs fresher information or heavier reasoning, the agent can escalate to a frontier cloud model. However, it asks first, and it screens for personal data before anything leaves the device. Code and tools run inside isolated sandboxes with controlled access to your files.

Most agent platforms ask you to assemble the parts yourself: an inference engine here, a harness there, a sandbox bolted on. Perplexity’s pitch is that the local build ships all of it in one installer.

The trade is obvious enough. You give up frontier model quality and gain privacy, repeatability and a flat cost. For document review and routine analysis that bargain looks reasonable. For research that needs the newest facts, it does not.

NVIDIA’s director of developer technology argued the timing has changed: “Local AI reached an inflection point.” Two years ago, he noted, running this at home was not super practical as hobbyist work.

Local AI has mostly meant a chat window talking to a model on your GPU. This is a different thing: a scheduler, a file system and a tool loop, all running where your data already lives. Therefore it lands closest to the people who could never send those files anywhere.

Law firms, clinics and finance teams have spent two years saying no to cloud agents for exactly that reason. Meanwhile, the hardware bar keeps the audience small for now. A 24GB card is still an expensive component, and that, more than the software, decides who gets to try this in 2026.

If local agents interest you, compare the leading agentic AI frameworks and what each one needs to run. And for a desktop that clears the video memory bar comfortably, read our HP OMEN 45L review.

Sources

NVIDIA Blog, VentureBeat, Unite.AI. *Photos: Palit GeForce RTX 5090 by PantheraLeo1359531, CC BY 4.0, cropped and resized; Gigabyte GeForce RTX 3090 Eagle OC 24G by PantheraLeo1359531, CC BY 4.0, cropped and resized.*