Perplexity Portable Computer Reaches Windows RTX PCs

The local agent has reached Windows

Perplexity Portable Computer is now available in the Perplexity app for Windows, three weeks after its initial Linux launch on NVIDIA DGX Spark. It is an agent rather than a conventional chat window: Perplexity says it can search local files, work with connected services, create documents, and run jobs on a schedule while its model and orchestration system operate on the PC.

The release matters because Windows gives the product a much larger potential audience than its original DGX-only launch. It also turns local inference into a feature inside a familiar consumer app. Users do not need to install a model server, select a quantization, or wire an agent framework to their files before they can try it.

What stays local, and what does not

NVIDIA says the Windows version uses a post-trained 27-billion-parameter model derived from Qwen and runs the model locally on the GPU. The surrounding planner, orchestrator, scheduler, and local search components also run on the device. Local jobs do not consume Perplexity Computer credits, so repeated file analysis or document work avoids per-token cloud charges.

That does not make every Portable Computer task offline. The agent can ask a cloud model to handle web research or work beyond the local model's capability. Perplexity says it requests permission before making that handoff. The approval prompt is the privacy boundary users should watch: anything approved for cloud processing has left the fully local path.

Connectors add another boundary. Outlook, OneDrive, Word, Google Drive, Gmail, Slack, and GitHub can give the agent useful working context, but those services are remote even when the reasoning model runs on your GPU. “Local model” describes the inference location. It does not automatically make every connected data source or resulting network request local.

The 24GB VRAM requirement is the expensive catch

Portable Computer for Windows requires a GeForce RTX or RTX Pro GPU with at least 24GB of video memory. That includes premium desktop cards and professional hardware, but excludes most gaming laptops and many mainstream RTX desktops. NVIDIA says support for DGX Station is expected later.

The hardware floor is understandable for a 27B model with room left for an agent runtime, yet it changes the audience. This is not a feature that makes an ordinary Windows laptop a private AI workstation. It is a simpler software layer for people who already own a high-memory NVIDIA GPU or were considering one.

The original Portable Computer launch was limited to Perplexity Pro and Max subscribers, while NVIDIA's Windows announcement directs eligible users to the desktop app. Prospective users should check the current plan requirement inside Perplexity before buying hardware solely for this feature.

How it compares with Perplexity Hybrid Compute

Portable Computer and Hybrid Compute solve related problems differently. Hybrid Compute on Apple Silicon divides a task between local and cloud models, automatically identifying potentially sensitive material before asking whether it can be sent away. Portable Computer starts from a local agent and escalates when a task needs the web or a more capable model.

For privacy, the useful question is not whether the product carries a “local” label. Check the task history for cloud escalation, review connector access, and deny the handoff when the files should remain on the PC. For cost, keep repeatable document and scheduling jobs local, where Perplexity says they use no Computer credits.

This launch is a meaningful usability step for local AI, but not a mass-market one. It packages a capable local agent into a consumer application while preserving an explicit cloud permission gate. The main compromise has moved from software setup to hardware price.