Perplexity Brings Its Portable Computer Agent
to Windows RTX PCs
Portable Computer, the fully local AI agent that previously shipped only on Linux, now runs on Windows PCs powered by NVIDIA RTX GPUs. Prompts and files never leave the device, and multi-step tasks can be completed entirely by a local model.
Finally expanding
to Windows RTX PCs
On September 14, 2026, Perplexity announced — jointly with NVIDIA — that it was bringing its fully local AI agent, Portable Computer, to Windows PCs. Details are available on the official NVIDIA blog and the official Perplexity blog.
Portable Computer is a fully local version of Perplexity's autonomous agent, Computer. While the regular Computer relies on cloud models to plan and execute multi-step tasks, Portable Computer runs that same plan-and-execute pipeline entirely with a model on the device. Prompts and the files it works on never get sent to the cloud, and any task completed locally does not consume Perplexity Computer cloud credits.
Portable Computer first shipped as a Linux release on August 25, 2026; this Windows release is its expansion. The idea of an agent that never touches the cloud, validated first on the narrower Linux audience, is now being brought to the much larger Windows RTX PC market.
Neither the prompt nor the files
ever leave the device.
Why the cloud
can be skipped
The same prompt produces a very different answer to "what leaves the device" depending on where it's processed.
The feature requires a Perplexity Pro or Max subscription (individual or enterprise), and it is distributed through Perplexity's Windows app (available via the Microsoft Store). The requirement is a compatible NVIDIA GeForce RTX PC or NVIDIA RTX PRO Workstation with 24GB or more of VRAM, which rules out most mainstream 8GB and 12GB-class gaming GPUs.
The local model running on Windows machines is "PPLX 27B" — a dedicated model that Perplexity post-trained itself, built on the Qwen 3.8 27B architecture and optimized specifically for NVIDIA RTX GPUs. Some early coverage of the announcement simply said "it runs Qwen 3.8 27B," but Perplexity's own product page clarifies this point: what ships on Windows RTX PCs is not raw Qwen 3.8 27B, but PPLX 27B, its post-trained derivative.
For tasks that need more reasoning power than the local model can provide, Portable Computer can escalate the work to Perplexity's cloud models. However, it is designed to explicitly ask the user for permission before sending any data off the device, so local-only processing remains the default. A built-in browser for web actions and Perplexity's sandbox environment, "SPACE," are also available locally.
Who it helps, and how
The benefit isn't uniform — what matters changes depending on your role.
Engineers
Agent workflows over proprietary code or sensitive customer documents can run locally, with no API token metering and no risk of cloud egress. On an RTX-equipped dev machine, even automating over a private repo never sends that data onto the network.
Business & PM (compliance)
In finance, healthcare, government, and other regulated industries where data residency rules can block cloud AI altogether, local execution lowers the adoption bar. The pitch to an auditor also gets much simpler: the data never leaves the device.
Heavy Computer users
Work completed locally does not consume Perplexity Computer cloud credits. The heavier your daily use of Computer, the more an investment in an RTX PC can cut ongoing costs.
Why local AI agents
matter right now
Until now, the default assumption for AI agents was that the prompt goes to the cloud and a cloud model does the processing. Most mainstream agent features, to varying degrees, still haven't broken out of that framing. What Portable Computer demonstrates is a different answer: the entire "plan, then execute" agent workflow can run on a local GPU.
Two trends underlie this. First, RTX-generation GPUs, from consumer cards up through workstation parts, have reached the point where their VRAM capacity and compute are enough to run large models locally. Second, as enterprise adoption of generative AI has grown, the constraint of "we can't hand this confidential data to a cloud AI vendor" has surfaced across industries. August's Linux launch read mainly as a technical proof of concept; this Windows rollout is what carries the idea into the mainstream PC market.
The next move, and the limits worth watching
There are three things an enterprise IT team can act on right away.
Pull procurement forward
Start evaluating purchases of RTX PRO workstations or high-end GeForce RTX machines with 24GB or more of VRAM, alongside mapping which workloads would use them.
Pilot on low-risk work first
Trial Portable Computer on internal document organization or report drafting before pointing it at your most sensitive primary-source material.
Set the escalation policy up front
Decide, before rollout, who gets to approve sending data to the cloud and under what conditions.
Optimism should still be tempered. The 24GB-or-more VRAM requirement is a high bar that excludes most mainstream consumer GPUs in the 8GB and 12GB class, so the benefit accrues only to users who have already made that hardware investment. PPLX 27B is also, at the end of the day, a 27-billion-parameter model — it doesn't match the reasoning power of frontier cloud models out of the gate, which is exactly why the escalation mechanism exists. If a "fully local" product ends up escalating to the cloud frequently in practice, its original privacy advantage erodes. How much work actually stays local in day-to-day use is something that will only become clear over time.