Perplexity's Local AI Agent Lands on Windows RTX PCs

Perplexity has brought Portable Computer, the fully on-device version of its agent platform, to Windows PCs with an NVIDIA GeForce RTX or RTX PRO GPU carrying at least 24GB of VRAM. The release landed on 14 September through the existing Perplexity app in the Microsoft Store, and is available to Pro and Max subscribers on both individual and enterprise plans.

The distinction worth holding onto is that this is not a chatbot running offline. Perplexity Computer is an agent that plans and executes multi-step tasks, and Portable Computer runs the whole stack locally — the model, the agent harness, the orchestrator and the scheduler all sit on the device. Scheduled recurring tasks keep firing without a round trip, and local Model Context Protocol servers let the agent reach desktop applications that never touch a network.

What runs on the machine

The default local model is Qwen 3.8 27B, post-trained by Perplexity to work with the Computer harness and optimised for RTX GPUs. Alongside it the company ships PPLX 27B, its own post-trained variant of the same base. NVIDIA's Nemotron 3.5 Lightning 30B is listed as coming soon, with Nemotron 3.5 ASR handling local dictation.

Connectors cover Microsoft Outlook, OneDrive and Word, Google Drive and Gmail, Slack and GitHub. Code and tool execution run inside isolated sandboxes with controlled access, including Perplexity's own SPACE sandbox — the same containment the cloud product uses. When a task genuinely needs stronger reasoning or live web information, the agent can reach Perplexity Search or any of 15-plus frontier models, but it asks permission before anything leaves the device.

The economics are the pitch

Work the agent finishes locally does not consume Perplexity Computer credits. That is the argument NVIDIA and Perplexity are both making: an agent that spends an hour grinding through a document review or an iterative analysis racks up a real bill against a metered API, and close to nothing against a GPU you already own.

"With Portable Computer, we really focused on just making this a really straightforward experience where you can get up and running very quickly," said Perplexity's vice president of engineering for infrastructure and enterprise. Nader Khalil, NVIDIA's director of developer technologies, framed the timing more broadly: local AI has "reached an inflection point," he said, and is "really killer for agents."

Where the local model actually sits

Perplexity's own numbers are candid about the ceiling. On its internal Local Knowledge Work Bench, PPLX 27B scores 85.4% against Qwen 3.8 27B's 82.6% — a real gain from post-training, on a benchmark the company built and grades itself. On harder reasoning, compact models still trail the frontier meaningfully. The company's measured compromise is hybrid escalation: handing the difficult step up to Claude Opus 5 recovered roughly 60% of the gap to frontier performance at about two-thirds the cost of running the whole task in the cloud.

The floor is the catch

A 24GB VRAM minimum means a GeForce RTX 3090 or newer, which excludes the overwhelming majority of consumer PCs and every laptop without a workstation-class card. Support for NVIDIA's DGX Station is listed as coming; Apple silicon does not appear on the roadmap at all. Portable Computer debuted in August on the DGX Spark desktop, a Grace Blackwell GB10 machine, and it was Linux-only until now.

So the honest framing is a hardware-gated preview of something that gets interesting later. The agent is real, the sandboxing is real, and the credits it does not burn are real — but the entry ticket is a GPU that costs more than most people's whole computer. Perplexity has built the software for a fleet of machines that mostly has not been bought yet.