Perplexity has brought Portable Computer to Windows: a version of its agent that runs on your machine instead of its servers.
What runs locally
Not just the model. Perplexity moved the model, the agent harness, the orchestrator and the scheduler onto the device. The agent plans and carries out multistep tasks the same way the cloud version does, but the work never leaves the machine.
The default local model is Qwen 3.8 27B, post-trained for Perplexity Computer and optimised for RTX hardware. Alongside it sit PPLX 27B, which Perplexity trained in-house, and NVIDIA Nemotron 3.5 Lightning, a 30-billion-parameter model that is still arriving.
What it costs and what it needs
The hardware requirement is the real gate: an NVIDIA GeForce RTX GPU or RTX PRO workstation with 24GB of VRAM or more. That is a high-end card, not a typical laptop. Access is limited to Pro and Max subscribers on individual and enterprise plans, through the existing Windows app in the Microsoft Store.
The payoff is that locally completed work does not consume Perplexity Computer credits. For anyone running an agent repeatedly on the same kind of task, that changes the arithmetic of using one at all.
Why it matters
Two things are converging here. Sensitive material stays on the device, which answers the objection that blocks agent adoption in a lot of regulated work. And the per-task cost drops to electricity, which removes the quiet anxiety of watching a credit balance while an agent iterates.
The catch is that both benefits are bought with a graphics card. This is not a cheaper way to use an agent — it is a different trade, capital up front instead of metered usage, and it only pays back at volume. Worth evaluating if you already own the hardware or your data cannot leave the building.