Perplexity and Nvidia Ship an AI Agent That Never Leaves Your GPU

Perplexity this week launched Portable Computer, a local edition of its agentic Computer platform built to keep models, data, and every step of a task on hardware the user already owns, starting with Nvidia’s DGX Spark and RTX-equipped Linux machines. Built jointly with Nvidia, the release bundles local models, an agent harness, an inference engine, and a security sandbox into one install rather than requiring a developer to wire those pieces together separately.

The stakes are about who pays for agents, and how much. Cloud agents that run for hours reviewing documents or checking their own work accumulate real API bills. Portable Computer’s pitch flips that: the identical workload, run on a GPU already purchased, adds nothing per token. That math changes the calculus for any team running agents continuously rather than in short bursts.

By default, nothing leaves the machine. A step only escalates to a cloud model after the system asks permission, and a PII classifier reviews what would leave the device before that call goes out. Nate, who runs infrastructure and enterprise engineering at Perplexity, said the team was chasing a “really straightforward experience where you can get up and running very quickly.” That is a pointed benchmark to set, given that DIY local stacks still ask developers to assemble weights, inference servers, and tool connectors themselves.

Perplexity’s research makes a structural argument: an agent harness built for frontier models that can absorb enormous contexts breaks down when it is pointed at small local models instead. The company’s own testing found Qwen 3.8 27B advertising a 260,000-token context window while its usable performance degrades past 100,000 tokens. The fix was to strip the harness down: a compact system prompt, a narrow core toolset, and MCP connectors rebuilt as lightweight command-line tools instead of token-heavy servers.

The benchmark numbers come from Perplexity, not an outside evaluator. Running Qwen 3.8 27B, Computer scored 82.6 percent on the company’s internal Local Knowledge Work Bench. That beat the open-source Pi harness, which scored 77.6 percent, and Hermes, at 74.0 percent, both tested on the same underlying model. Swapping in Perplexity’s own post-trained PPLX 27B model lifted the score further, to 85.4 percent. No independent third party’s benchmark accompanies the release.

More commercially telling is what happens when the system escalates. On the coding benchmark Terminal Bench 2.1, the local Qwen model alone reached 59.6 percent at essentially no cost. Handing off the hardest steps to a Claude Opus 5 advisor in the cloud pushed that score to 73.0 percent, at roughly $0.415 per task; letting the cloud model handle everything scored 82.4 percent, at $0.65 per task. Selective escalation closes most of the capability gap for about two-thirds of the frontier price, and the user decides when that spend happens.

The constraints are concrete, not cosmetic. Linux is required, and so is a GPU carrying 24GB of VRAM or more, a bar roughly equivalent to a GeForce RTX 3090, which rules out most consumer machines by default. The companies say Windows will follow next month; Apple silicon, where a great deal of local AI tinkering already lives, does not appear on the roadmap at all. Nvidia’s Nader positioned tools like Ollama as fine for simply getting a model running, but inadequate once the workload turns agentic, casting Portable Computer as a layer sitting above existing local inference tools rather than a replacement for them.

For Nvidia, the launch builds a case for buying DGX Spark hardware instead of renting cloud GPU capacity, extending a partnership with Perplexity that traces back to a 2025 sovereign AI collaboration. For Perplexity, which faces legal pressure from publishers, including a lawsuit from The New York Times, a product whose economics do not hinge on metering tokens opens a route into privacy-sensitive sectors such as law, healthcare, and finance. Engineering teams sizing agent infrastructure spend for 2027 budgets should treat the Terminal Bench escalation figures as a starting point for their own cost modeling rather than a finished answer, given that Perplexity has not published independent verification of the results.

Reported by VentureBeat on 25 August 2026.