Perplexity launches Portable Computer, an AI agent that runs entirely on your own machine
Built with Nvidia, the new product lets an AI agent handle hours of document work without sending your files to the cloud or charging you per task.
Key points
- Perplexity launched Portable Computer on Monday in partnership with Nvidia, bringing its AI agent software to local hardware for the first time.
- The product runs on Nvidia DGX Spark desktop machines and Linux PCs with compatible Nvidia RTX graphics cards, with Windows support coming in September 2025.
- All processing happens on the user's own device by default, with the system asking permission before escalating any step to a cloud model.
- On Perplexity's own 53-task benchmark, its co-designed agent scored 85.4% accuracy versus 77.6% for the nearest open-source rival.
- The minimum hardware requirement is an Nvidia RTX graphics card with at least 24 GB of video memory, roughly a GeForce RTX 3090 or newer.
Perplexity, the AI search company, has released a product it calls Portable Computer: a version of its "Computer" agent platform that runs entirely on hardware you already own. An AI agent is software that carries out multi-step tasks on its own, reading documents, flagging problems, writing a summary report. Until now, that kind of extended task meant sending files to a remote server and paying for every step.
Portable Computer changes both of those things.
What does it actually do?
The software packages everything an AI agent needs into a single download: the AI model, the tools it uses, and the orchestration logic that keeps it on track. Install it, point it at your documents, and it works.
At a press briefing Monday, Perplexity's vice president of engineering, Nate, showed the agent reviewing a folder of 1099 tax forms and investment statements on a local machine. It flagged cases where a hypothetical investor was paying unnecessary fees. The billing counter on screen stayed at zero because nothing left the device.
A second demo showed the hybrid angle. An agent analysed a spreadsheet of sales funnel data locally, then pushed the finished report to a Slack channel. Local-first doesn't mean cut off: the system also connects to Google Drive and GitHub, and can hand a task to a more powerful cloud model if the local one hits its limits, but only after asking permission.
Why does running it locally matter?
Privacy is the obvious reason. Financial records and internal business data are exactly the kind of material many people are uneasy uploading to a third-party server. Here, those files never leave your desk.
Cost is the other reason, and it compounds fast. A simple chat message takes a second. An agent that spends two hours combing through hundreds of documents, checking its own work, and producing a report can generate an enormous number of individual computation steps, each carrying a small fee on a cloud service. On your own machine, that marginal cost drops to near zero.
"With agents, you want these agents always on if you can," said Nader, Nvidia's director of developer technology, at the briefing. "You were not metered by the token. You were not paying for the token. So it's really killer for agents."
How does it perform against the competition?
Perplexity published a research paper alongside the launch arguing that small local AI models need a purpose-built scaffolding of instructions and tools around the model, rather than the general-purpose frameworks designed for much larger cloud models. The company says it co-designed that scaffolding and model together, and the numbers reflect it.
As we reported on 21 August, Alibaba's Qwen 3.8 27B had already matched cloud rivals on coding tests before Perplexity built its product around the same model.
| System | Local Knowledge Work Bench | BrowseComp web research |
|---|---|---|
| Perplexity Computer + PPLX 27B | 85.4% | |
| Perplexity Computer + Qwen 3.8 27B | 82.6% | 66.7% |
| Pi scaffolding + Qwen 3.8 27B | 77.6% | 50.2% |
| Hermes scaffolding + Qwen 3.8 27B | 74.0% | 43.9% |
These figures come from Perplexity's own internal benchmark of 53 tasks, which the company says it will open-source. Independent verification doesn't yet exist, so treat the gaps as directionally interesting rather than settled fact.
At launch, users can choose between Qwen 3.8 27B, an open-source model from Alibaba's research lab, and PPLX 27B, Perplexity's own fine-tuned version, with a model from Nvidia coming soon.
Who can use it right now?
Portable Computer is available today on Linux for subscribers on Perplexity's Pro, Max, Enterprise Pro, and Enterprise Max plans. Windows support follows in September. The minimum requirement is an Nvidia RTX card with 24 GB of video memory, roughly a GeForce RTX 3090 or newer. That's a high bar for a home PC but standard on many professional workstations.
First reported by VentureBeat, this launch is one of the more serious attempts to move genuine AI workload off remote servers and onto local machines. The honest caveat: the benchmark is Perplexity's own, the hardware bar is steep, and the Windows version isn't here yet. What to watch is whether independent testers close or widen those accuracy gaps once they get the tool in their hands.
