Perplexity Portable Computer: a local agent at zero token cost

Perplexity shipped Portable Computer today, a version of its agent that runs inference entirely on hardware you own. The pitch is that work done locally burns no billing credits, so the marginal cost of a token approaches zero. That is a real change in how agent economics work — and it comes with a hardware floor and a subscription gate that decide whether it applies to you.

What actually happened

Per VentureBeat's report, Portable Computer runs on Nvidia DGX Spark and on Linux machines with Nvidia RTX GPUs. The stated minimum is an RTX GPU with 24GB of VRAM — a GeForce RTX 3090 or newer — and the setup scales across multiple connected devices for larger models.

The model lineup is Qwen 3.8 27B, PPLX 27B (Perplexity's own post-trained variant), with Nvidia Nemotron 3.5 Lightning listed as coming soon. Linux availability starts today. Windows follows in September 2026. Access is limited to Pro, Max, Enterprise Pro, and Enterprise Max subscribers.

"Zero token cost" means what it says and no more: inference that completes on your machine does not draw down credits. The subscription that unlocks the feature is still a subscription.

Why zero token cost matters for your business

Metered inference kills the boring high-volume jobs first. Reclassifying 40,000 SKUs, summarizing every support thread from last quarter, re-tagging an image library — these die in planning because someone multiplies tokens by rate and flinches. When the marginal token is free, the jobs that were never worth the API bill become worth running overnight. That is the actual unlock, and it has nothing to do with chat.

27B is not frontier, and that is usually fine. A 27B-class model is competent at classification, extraction, summarization, and routing. It is not the thing you want writing your contract language. Route by task: local for volume, cloud for the handful of calls where quality is the whole point. Anyone who tells you to pick one is selling you something.

24GB of VRAM is the real price. An RTX 3090 or newer is the entry ticket, and DGX Spark is a different budget entirely. Before this reads as "free inference," price the hardware and the electricity, then divide by the workload you would actually put on it. It is the same break-even that Apple's 512GB Mac Studio poses from the other direction — the industry is converging on local, and the math still has to clear.

Keep the harness yours. The thing worth building is not "we use Perplexity locally." It is a pipeline where the model endpoint is a config value. Qwen 3.8 27B today, Nemotron 3.5 Lightning next month, something else in six months — none of that should touch your agent logic.

Sources: VentureBeat.