Perplexity’s Portable Computer Runs Agents Locally—but Needs a 24GB Nvidia GPU

Perplexity has launched Portable Computer, a local-first edition of its Computer agent platform for Linux machines with compatible Nvidia hardware. The promise is not a permanently offline agent: work begins on the device, but the system can ask to send an individual step to a more capable cloud model. That makes the product’s central tradeoff unusually explicit—keep files, models and agent execution local when possible, then let the user decide when stronger remote help is worth it.

Portable Computer is meant to solve a setup problem that sits beyond simply downloading an open model. It bundles a local model, inference engine, agent harness, tools, app connectors and a security sandbox into one application. The harness is the orchestration layer that gives a model instructions, tools and a sequence for completing multistep work; the sandbox isolates its code and tool execution.

That permission step is the meaningful privacy mechanism. Before an escalation, Portable Computer classifies personally identifiable information in the outgoing context and shows the user the data proposed for transfer. Perplexity says the remote model returns text guidance only, rather than receiving access to local files or tools. The system can still work with external services: its locally executed workflows connect to Slack, Google Drive, Gmail and GitHub.

The first audience is narrow

The initial release is available to Perplexity Pro, Max, Enterprise Pro and Enterprise Max subscribers on Linux. It requires an Nvidia RTX GPU with at least 24GB of video memory, roughly the GeForce RTX 3090 class or newer, or a DGX Spark. That requirement puts Portable Computer beyond most consumer PCs, even for people already paying for an eligible Perplexity plan.

What ships, and what is planned

• Users can configure Qwen 3.8 27B or Perplexity’s post-trained PPLX 27B model at launch.
• Nvidia Nemotron 3.5 Lightning is planned for a later release.
• Windows support is expected in September; macOS support has not been announced.

Perplexity is betting the stack matters as much as the model

Perplexity’s case is that smaller local models need a more constrained operating layer than frontier cloud models. The company says it built a minimal harness with a short system prompt, a limited core tool set and skills that load only when needed. It also converted some connectors, including Gmail and GitHub, into command-line tools and says its harness will not operate if OS-level sandboxing is unavailable.

The performance evidence so far is company-reported. Perplexity says its Computer harness running Qwen 3.8 27B scored 82.6% on its internal 53-task Local Knowledge Work Bench, versus 77.6% for Pi and 74.0% for Hermes using the same model; its PPLX 27B reached 85.4%. Those results test Perplexity’s own benchmark, which the company says it plans to open-source, so they are an early signal rather than an independent verdict on the product’s advantage.

Perplexity also reports the limits of the local approach on Terminal Bench 2.1. Fully local Qwen scored 59.6%; escalating to a Claude Opus 5 advisor raised that to 73.0% at an estimated $0.415 per task, while the frontier model alone reached 82.4% at $0.65. Portable Computer’s real test is therefore not whether it eliminates cloud AI, but whether its permissioned hybrid design gives high-end Linux users enough local capability to reserve cloud calls for the work that actually needs them.