Perplexity Brings Its Portable Computer Agent to Windows PCs with NVIDIA RTX

Perplexity has expanded its local AI offering to Windows, making Portable Computer available through the Perplexity app on compatible NVIDIA GeForce RTX PCs and NVIDIA RTX PRO Workstations. The agent is designed to plan and execute multistep tasks on the device, while allowing users to approve cloud assistance when a job needs more advanced reasoning.

The release matters because it moves Perplexity’s agent from a primarily cloud-oriented experience toward a hybrid model that can work with sensitive files locally. According to NVIDIA, locally completed tasks do not consume Perplexity Computer credits, and information can remain on the PC unless the user gives permission to send it off-device.

What Perplexity Portable Computer does locally

Portable Computer is described as a local version of Perplexity Computer, the company’s agent for carrying out multistep work. It can analyze files, combine information from different sources and handle recurring tasks using local models accelerated by NVIDIA GPUs.

The Windows release builds on earlier support for NVIDIA DGX Spark systems and RTX PCs running Linux. It is available for NVIDIA GeForce RTX and RTX PRO GPUs with at least 24GB of VRAM. NVIDIA says support for NVIDIA DGX Station is expected later, but it did not provide a specific timetable.

The product combines local processing with Perplexity Computer’s existing tools, including a built-in browser and the company’s SPACE sandbox. It can also connect to services including Microsoft Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub. Those integrations are intended to let the agent operate across files and workplace applications rather than only answer questions in a chat window.

A local model such as Qwen 3.8 27B is included in the setup. NVIDIA says the model has been post-trained for Perplexity Computer and optimized for NVIDIA RTX hardware. The stated goal is to reduce the configuration burden normally associated with running local AI models, where users may otherwise need to select models, install runtimes and tune a software stack.

A hybrid path for sensitive work

The main product distinction is the division between work that can remain on-device and tasks that may benefit from cloud models. Portable Computer can identify when a request requires additional cloud-based reasoning, but NVIDIA says it asks the user for permission before sending information off the PC.

That design could be relevant for teams handling financial records, internal documents, source code or customer information. In NVIDIA’s examples, the agent reviews brokerage summaries, tax returns and other financial documents locally, with figures linked to the exact file and page. In another example, it reviews open pull requests in a connected GitHub project, organizes them by status and suggests documentation updates for review.

NVIDIA also describes a startup workflow in which Portable Computer examines a funnel export to identify where users drop out between installation and their first completed task, then posts findings to a team’s Slack channel. These are vendor-provided examples rather than independently verified demonstrations, and the source does not disclose how reliably the agent performs each workflow in production.

The local-cloud boundary also introduces an important operational question: users and organizations will need to understand exactly which files, excerpts or tool outputs leave the device when cloud support is approved. The permission step is a control, but it does not by itself establish the full data-governance behavior enterprises may require.

What the evidence shows—and does not show

The availability announcement comes from the NVIDIA Blog, making NVIDIA the primary source for the product, hardware and integration details. Perplexity’s own independent announcement, pricing information, Windows system requirements beyond the stated VRAM threshold and performance measurements are not included in the supplied evidence.

NVIDIA’s description confirms the supported hardware category, the local model setup, the available connectors and the ability to escalate selected work to cloud models. It does not provide latency figures, throughput comparisons, task-success rates or a benchmark against running the same agent entirely in the cloud.

The article also makes no independently verified adoption claim. References to broader access for Windows users describe the scope of the release, not the number of customers using it. Any claims about privacy, cost savings or productivity should therefore be treated as product positioning until tested across representative workloads.

The 24GB VRAM requirement is especially significant for buyers. It places the product outside the reach of many everyday Windows laptops and desktops, while making it more relevant to high-end gaming PCs, professional workstations and organizations already investing in local GPU infrastructure. The hardware requirement may help Perplexity deliver a more capable local experience, but it limits the addressable installed base.

Implications for builders and enterprise buyers

For AI builders, Portable Computer illustrates a practical architecture for agent deployment: keep document retrieval and routine analysis on the endpoint, then use cloud models selectively for harder reasoning. That approach can reduce unnecessary data transfer and may lower cloud usage for repeatable tasks, although the source does not quantify savings.

Product teams evaluating the system should test more than model quality. They will need to examine connector permissions, citation accuracy, failure recovery, audit logs, approval flows and the behavior of actions that change external systems. Reviewing a GitHub queue is materially different from submitting a code change, sending an email or posting an analysis to Slack. The supplied announcement confirms integrations, but not the limits or safeguards for every action.

For enterprises, the local option could be attractive where data residency or confidentiality makes a fully cloud-hosted agent difficult to approve. At the same time, local deployment shifts responsibility toward endpoint management, GPU availability, model updates and device security. A company may gain control over sensitive files while taking on a more complex fleet and software-support problem.

The launch also gives NVIDIA another route to make RTX hardware relevant to AI workloads beyond model experimentation. If Perplexity’s agent can turn local GPUs into a usable business tool without extensive configuration, hardware buyers may evaluate workstation capacity through agent workloads rather than graphics performance alone. That conclusion remains to be demonstrated through independent testing.

What to watch next

The first signal will be real-world testing on Windows systems with different RTX GPUs and 24GB-plus memory configurations. Reviewers and enterprise pilots should measure task completion, latency, VRAM consumption and behavior when the agent switches from local processing to cloud models.

Perplexity’s documentation should also clarify how users approve cloud escalation, what data is transmitted, how local files are indexed and whether organizations can enforce policies centrally. Support for NVIDIA DGX Station is another concrete milestone, particularly for teams considering dedicated local AI systems.

Finally, the market will need evidence on whether connectors such as Microsoft Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub support reliable workflows at scale. The difference between a compelling demonstration and a dependable enterprise agent will be visible in error handling, permissions and reviewable outputs.

Creati.ai perspective

Perplexity Portable Computer is a notable deployment step, not proof that local agents have solved the hardest reliability and governance problems. Its strongest proposition is architectural: sensitive, repetitive work can stay on a capable Windows device while users retain an explicit path to cloud reasoning.

The release will matter most if Perplexity can show that this boundary is understandable, enforceable and useful in daily workflows. For now, the hardware threshold and the NVIDIA-controlled evidence make this a promising option for well-equipped builders and workstation users, rather than a broadly accessible replacement for cloud agents.