Perplexity’s new agent runs entirely on your GPU — with one expensive catch
The New Stack Amanda Caswell ● Covered by 2 sources
Perplexity’s local Computer agent is now in the Windows app. It runs on your RTX GPU, but only if you’ve got 24GB of VRAM or more.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Perplexity has brought its local Computer agent to Windows, and the pitch is simple: keep the work on your own machine when you can, push it to the cloud when you must. The catch is just as simple. On Windows, Portable Computer only runs on compatible Nvidia GeForce RTX and RTX PRO GPUs with at least 24GB of VRAM.
That puts it in a different class from the usual local-model tools. Perplexity isn’t just letting Windows users run a model; it’s packaging model runtime, orchestration, security, and hardware integration into a single agent experience. On this platform it supports PPLX 27B, Perplexity’s post-trained model, and Qwen 3.8 27B, both tuned for RTX GPUs. It also comes with a built-in browser, tool calling, and Perplexity’s own SPACE sandbox.
The local-first part matters, but only up to a point. Perplexity ships connectors for Microsoft Outlook, OneDrive, and Word, plus Google Drive, Gmail, Slack, and GitHub, which means “local” doesn’t mean sealed off from the outside world. Tasks done on the machine can process files without sending documents to a cloud model. But once an agent can touch local files and remote APIs at the same time, the hard part becomes deciding what it actually needs to reach.
That’s why the hybrid fallback matters so much. If the local model decides it needs more reasoning power, it can hand the task off to Perplexity’s cloud models. Nvidia says the agent checks whether cloud support would help and asks the user for permission before any data leaves the machine. For sensitive or regulated work, that split is the whole point. Local processing can chew through source code or financial records without feeding everything into a hosted model.
Perplexity is also trying to make the economics friendlier. Work finished locally doesn’t use Perplexity Computer credits, which is a neat incentive if your GPU is up to the job. The service is available with Perplexity Pro at $20 a month and Max at $200 a month, across individual and enterprise plans, with Nvidia DGX Station support coming later.
My take — AI-written commentary, not fact-checked reporting
This is the right direction, and also a very Nvidia-shaped gatekeeping exercise. Local agents only become serious when they can actually do work without sending every file to a cloud somewhere, but 24GB of VRAM keeps the club small for now. The industry keeps calling everything an agent; the unglamorous part is still GPUs, sandboxes, connectors, and permission prompts.
Read more about this at: The New Stack