Perplexity brings local AI agents to Windows RTX PCs

Sending your company's raw financials, sensitive tax returns, or private codebase to a cloud AI has always felt like a necessary gamble. Building a local alternative, on the other hand, usually involves hours of driver troubleshooting and python environment hell.

Perplexity just solved that tradeoff for high-end PC owners. The search and agent platform officially brought its Portable Computer agent to Windows PCs powered by NVIDIA GeForce RTX and RTX PRO GPUs.

There is one catch: you need at least 24GB of VRAM to run it. The release expands on previous support for Linux RTX rigs and NVIDIA DGX Spark hardware, with DGX Station support promised soon.

Local execution with a hybrid escape hatch

Portable Computer is essentially a local counterpart to Perplexity Computer, built to plan and execute multi-step workflows directly on your machine.

Two big perks stand out immediately. Everything sensitive stays strictly on-device, and anything solved locally does not chew through your paid Perplexity cloud credits.

Instead of making users tinker with complex local software stacks, the Windows app ships with pre-configured models like Qwen 3.8 27B, post-trained specifically for agent tasks and tuned for RTX silicon.

It also hooks right into your actual daily toolchain. The agent features a built-in browser, Perplexity's proprietary SPACE sandbox, and native connectors for Microsoft Outlook, OneDrive, Word, Google Drive, Gmail, Slack, and GitHub.

That means an engineer can point it at a local repo to triage open pull requests and flag stale documentation. A founder can have it analyze activation drop-offs from raw user funnels and post findings to Slack without leaking customer records.

When a task exceeds local reasoning capabilities, Portable Computer identifies the bottleneck and asks for explicit user permission before routing the prompt to more powerful frontier models in the cloud.

Why it matters

For power users, developers, and finance pros handling strictly confidential records, this turns local agentic workflows into a genuine one-click reality. You get automated file digestion and workspace actions without paying per-token API charges or risking data leaks.

We think Perplexity and NVIDIA found the smartest compromise in modern AI deployment. Pure cloud agents scare compliance departments, while standalone local models frequently hit reasoning dead ends. A local-first workflow backed by a permission-based cloud safety valve feels like the exact blueprint enterprise desktop AI needs.

The obvious bottleneck is hardware. Capping compatibility at 24GB VRAM restricts this to workstation cards and top-tier GPUs, keeping mainstream laptop users on the sidelines for now. But as smaller open models gain ground, this release sets a clear standard for how local and cloud compute should share the desk.