Models & agents

Perplexity brings Portable Computer to Windows, needing an Nvidia RTX GPU with at least 24GB of VRAM

September 14, 2026 at 4:35 PM PT

Nvidia promotional image for Perplexity Portable Computer on Windows PCs powered by RTX GPUs

Image: NVIDIA

Why it mattersA local agent with a real GitHub and file-system connector changes what "private code review" looks like on a workstation, but a 24GB VRAM floor keeps this out of the reach of most developer laptops.

Perplexity has added Portable Computer to its Windows app on 14 September, letting the agent run locally on Nvidia GeForce RTX and RTX PRO GPUs. Nvidia states in the joint announcement that the card needs at least 24GB of VRAM. The New Stack, covering the launch the same day, calls the VRAM floor the "one expensive catch".

Portable Computer is the local version of Perplexity Computer, the company's browser-and-tool-calling agent. Nvidia writes that the app bundles orchestration, a built-in browser, tool calling, and Perplexity's SPACE sandbox alongside a post-trained local model. The New Stack reports that the Windows build supports two models: Perplexity's own PPLX 27B and Qwen 3.8 27B, both optimised for RTX GPUs.

What the local run buys

Nvidia's post gives three worked examples of what the local mode does that a cloud agent does not. For engineering, it can review open pull requests in a connected GitHub project, sort them by status, and flag outdated documentation with proposed updates for review. For finance, Nvidia writes, an agent can point at two years of brokerage summaries and tax returns and trace the holdings creating the most fee and tax drag, "with every figure cited to the exact file and page, all without a document ever reaching a chatbot". For a startup, it can pull a funnel export locally to find where signups drop off, then post the top findings to Slack.

The New Stack notes that Perplexity ships connectors for Microsoft Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub, and that this makes the "local" label narrower than it sounds. Sensitive files can stay on the machine, but the agent still reaches external services when it needs to. Nvidia writes that when a task calls for more reasoning than the local model can provide, Portable Computer identifies the escalation and asks the user before sending anything to Perplexity's cloud.

The hardware bar and the pricing

The 24GB VRAM requirement rules out every consumer Nvidia card below the RTX 4090 and the RTX 5090, along with most laptop GPUs. Nvidia writes that DGX Station support is coming later, on top of the existing Linux and DGX Spark builds. This is the third platform launch in under three weeks: Portable Computer shipped on Linux and DGX Spark on 25 August, and hybrid compute for Apple silicon followed a week later.

The New Stack reports that Portable Computer is included with Perplexity Pro at $20 a month and Max at $200 a month, on both individual and enterprise plans, and that tasks completed locally do not draw down Perplexity Computer credits. Those figures come from The New Stack's reporting and Perplexity's plan pages, and are the vendor's own list prices.

For teams handling regulated data on a workstation with the right card, a local agent that can read a codebase and call the GitHub API without shipping the source elsewhere is a different tool from a cloud agent with a promise about data handling. For everybody else, the hardware bar means this stays a workstation feature rather than a laptop one for now.

Source

Primary source: Perplexity Portable Computer Is Now Available on Windows, Powered by NVIDIA RTX, NVIDIA Blog, 14 September 2026. Reporting: Perplexity's new agent runs entirely on your GPU, with one expensive catch, The New Stack.

Source: NVIDIA Blog

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

More from AI News

Andon Labs opens Pion, its platform for running agents against real businesses, as a research preview

Andon Labs, the Y Combinator lab behind the Vending-Bench evaluation, has opened Pion as a research preview, a platform that gives one persistent agent the tools to run a real business: email, phone, banking, a browser and secure computing environments.

Source: Hacker NewsModels & agents

Entelligence benchmarks GPT-5.6 Luna against GPT-6 Astra on code review, and reports 74% precision at 28 times lower cost

Entelligence AI reran its code-review benchmark with the cheapest OpenAI model against the most expensive one on 50 public pull requests, and reports that GPT-5.6 Luna found three-quarters of the bugs GPT-6 Astra found for less than 4% of the money, with the gap concentrated on authentication and permission code.

Source: Hacker NewsModels & agents

Zeeshan Yaseen ran two GEO experiments, and third-party listicles produced 85.8 percent of AI citations

Two hand-tracked GEO experiments by consultant Zeeshan Yaseen found third-party listicles produced 85.8 percent of AI citations, with owned listicles at 14 percent.

Source: PressGo-to-market