Perplexity Portable Computer Moves Its AI Agent onto Local PCs

Perplexity puts its agent on your own hardware

For most of 2026, Perplexity has pitched its Computer agent as a cloud service. You describe a result you want, and the system plans the work and runs it on Perplexity's servers. On August 25 the company changed direction. It launched Portable Computer, a version of the agent built with Nvidia that runs on hardware the user already owns. The first supported machines were Nvidia's DGX Spark desktop system and Linux PCs with RTX graphics cards.2 By default, the model, the user's files and the work itself stay on the machine. Any work finished locally uses none of the billing credits that cloud Computer consumes.2

The rollout has moved quickly since then. Windows support arrived in mid-September.46 By late September Perplexity had also announced support for AMD Ryzen AI Max systems, so the product no longer depends on Nvidia hardware alone.17 Perplexity's product page now covers both Windows and Linux machines that have at least 24GB of video memory.3

What actually runs locally

The headline isn't that a 27-billion-parameter model fits on a desktop. Hobbyists have been running models that size locally for some time. What's new is how much of the agent runs on the device. Coverage agrees that the orchestrator, planner, tool router, scheduler, durable task queue and local search index all run on the machine.45 In practice, the parts that decide what to do next and keep long jobs running are no longer hosted by Perplexity.5 One analysis described the result as closer to an agent operating environment on the desktop than a local chatbot.5

Two local models were available at launch. One is the open-source Qwen 3.8 27B. The other is PPLX 27B, a version Perplexity post-trained to work with its own harness. Nvidia's Nemotron 3.5 Lightning is listed as coming soon.2 One commentator pointed to an interesting detail in that post-training: the model was taught to finish as much of a task locally as it can and to spot when a step really needs the cloud. That makes it a model trained around cost as well as capability.11

The engineering approach goes the same way. Perplexity found that Qwen 3.8 27B advertises a context window of about 260,000 tokens but starts to struggle past roughly 100,000. So the company built a deliberately small harness: a short system prompt, a handful of core tools, and extra capabilities that load as "skills" only when needed.2 Perplexity's argument is that small local models do badly inside general-purpose agent frameworks built for frontier models, and that the model and harness have to be designed together.18

The company's own benchmarks support that argument. On its 53-task Local Knowledge Work Bench, the local stack reportedly scored 82.6% with Qwen 3.8 27B and 85.4% with PPLX 27B, against 74.0% for the open-source Hermes harness running the same model.18 Perplexity also says it matched or beat the Pi and Hermes harnesses on accuracy and used the fewest tokens across its tests.9 All of these numbers come from Perplexity, and at least one analyst has said they need independent testing.11 That's a fair warning. Benchmarks a company designs for its own product tend to favor that product.

Local-first, not local-only

The coverage agrees that Portable Computer isn't an offline tool. Perplexity calls it "local-first." Every task starts on the device, and the system asks permission before sending any step to a stronger cloud model.2 According to the product page, more than 15 cloud models are available for those escalations.3 The agent also connects to Google Drive, Gmail and GitHub.2 The Windows version adds Outlook, OneDrive, Word and Slack, plus local Model Context Protocol servers so the agent can use other desktop apps.4

Perplexity frames the result as private and cheap by design. In its technical write-up, it says sensitive data doesn't leave the device without permission and local models carry no inference fee.1 Code and tool execution run in isolated sandboxes with limited access to files and connected apps.5 The AMD release adds a sensitive-content classifier to the package.17 That matters, because a local agent with broad file access isn't automatically safer than a cloud chatbot. It may be able to reach more valuable data.5

Where the reporting diverges

On the facts, outlets mostly agree. The differences are in emphasis and in a few details that changed during the rollout.

The first is timing. Most outlets date the Windows release to September 14, the day Nvidia announced it and Perplexity CEO Aravind Srinivas posted about it on X.46 One detailed timeline instead dates the expansion to September 21, which coincides with a wider Computer update.15 The likeliest explanation is a staged release. TechRepublic's September 21 piece described Windows availability as settled by then.14

The second is which models run on Windows. Nvidia's Windows materials used Qwen 3.8 27B as the example local model.46 TechRepublic, citing Perplexity's documentation, said PPLX 27B was the only local model on Windows and that the stock Qwen model wasn't yet supported on Windows RTX PCs.14 The product page backs the narrower version: it tells Windows users to set up PPLX 27B and notes that only one local model runs at a time.3

The third is tone. ZDNET called the launch a "game changer" for speed, security and cost, while noting the strict requirements.1 VentureBeat called it one of the most aggressive attempts yet to move serious agent work off the cloud.2 Tom's Hardware focused on the hardware limit, pointing out that the 24GB requirement depends on memory capacity rather than GPU generation and applies across Nvidia's consumer range.6 Other analysts were cooler. They said the 24GB threshold puts the product in high-end and workstation territory and keeps it well away from the average Windows PC.79

Why Perplexity is doing this

Portable Computer fits a broader pattern. Perplexity has spent 2026 turning its search product into a family of agent products. Computer launched in February. A Personal Computer extension for local devices followed in April.12 Personal Computer reached Windows in July, and Hybrid Compute arrived on Mac in September.15 At launch, the cloud Computer relied on Anthropic's Claude Opus 4.6 as its main reasoning engine and spread work across 19 models.9 That design means paying other companies for inference on every task, and the costs grow as agents run longer jobs.

I read Portable Computer as partly a response to that cost structure. When work runs on hardware the customer has already bought, Perplexity's marginal cost for that work drops toward zero. The company can then market "unmetered" use, a word Srinivas used himself when announcing the Windows release.6 Cloud Computer, by contrast, is metered: Perplexity prices credits at 100 to the dollar.16 One analysis described the change as moving from metered inference to compute the user owns.7 In that model, frontier cloud models become something the local agent calls when it needs to, not the place where all the work happens.7

Hardware partnerships are the other half of the story. BigGo reported that the launch coincided with reports that Nvidia might invest as much as $30 billion in Perplexity.9 Those reports haven't been confirmed, but a product built to sell high-memory GPUs would suit such a relationship well. The AMD expansion suggests Perplexity wants to be the agent layer on any capable local machine, not one chipmaker's showcase.17 HP ZBook Ultra G3a laptops are also expected to ship with Perplexity preinstalled in October.15

Who this is for, for now

Access is limited. Portable Computer requires a Pro, Max, Enterprise Pro or Enterprise Max subscription, so free users can't use it.9 On top of that, users need a GPU with 24GB of video memory, roughly a GeForce RTX 3090 or newer. Perplexity's vice president of engineering called that "sort of the floor" for a good experience.2 The realistic early users are professionals handling sensitive files, such as code under NDA, client records or unreleased financials, who already own or are willing to buy suitable hardware.11 Perplexity's own example points the same way. It describes a logistics controller who schedules the agent to check freight invoices against carrier rate sheets each morning, without the data leaving the company's device.4

The bigger question is whether a local agent becomes the normal way to deliver Perplexity's product. As local models need less memory, many more machines will be able to run it, and the product will matter more.7 For now, Portable Computer is a working, packaged answer to a hard problem in agentic AI: who pays for the compute, and where the data goes. It's still a premium option for people with high-end PCs, but it shows that Perplexity wants its agent to run wherever the customer's hardware is.