Running Perplexity's agent locally costs $4,800

Perplexity and Nvidia have brought Computer, Perplexity's agentic assistant, to the desktop as Portable Computer, an agent that runs entirely on local models on hardware starting at a $4,800 DGX Spark. Perplexity announced the release on Threads; the hardware requirements and pricing were detailed by The New Stack.

At a glance

• A deterministic orchestrator, rather than a second model, assembles context, enforces policy and executes approved tool calls inside an OS-level sandbox that limits processes, filesystem paths and network access.
• On Perplexity's internal 53-task Local Knowledge Work Bench, Computer scored 82.6% against 77.6% for Pi and 74% for Hermes, all on the same Qwen3.8-27B base model and DGX Spark hardware.
• Access is limited to Pro, Max, Enterprise Pro and Enterprise Max subscribers, and locally completed tasks consume no usage-based credits unless the agent has to call Perplexity's cloud models.

Local agents have mostly been a do-it-yourself exercise: a model server, a tool layer and a sandbox, assembled by hand. Portable Computer packages that stack, but the price of entry keeps it inside a narrow group of buyers with Nvidia hardware already on the desk. The benchmark numbers also read as an argument about harnesses rather than models, since the comparison holds the base model fixed.

The cheapest supported setup is a $4,800 DGX Spark

Two hardware paths are supported at launch. The first is an Nvidia DGX Spark desktop running DGX OS; the second is a conventional PC running Ubuntu on ARM or x64 hardware with an Nvidia RTX card carrying at least 24GB of GPU VRAM.

The New Stack puts the DGX Spark at $4,800 and notes that even an older RTX 3090 card with 24GB of VRAM currently costs well over $1,500, with memory prices unlikely to fall in the near term.

At launch, users choose between Qwen3.8-27B and PPLX 27B, which Perplexity describes as a new post-trained model built on it. Nvidia's Nemotron 3.5 Lightning is due later, and Portable Computer also accepts outside models running on a user's own inference server.

Qwen3.8-27B supports 260,000 tokens but starts to struggle past 100,000

Nate Kupp, Perplexity's vice president of Computer Enterprise and Infrastructure, told The New Stack that the team had to revisit almost everything throughout the stack. Many of Computer's capabilities were reused, but the harness and model configuration were rebuilt for local hardware, and the harness accounted for most of the engineering work.

Perplexity says Qwen3.8-27B supports a 260,000-token context window but begins to struggle beyond 100,000 tokens. The harness therefore keeps the core prompt and tool set small, loads additional skills only as needed, and exposes commonly used connectors as command-line tools instead of full Model Context Protocol definitions.

Every task starts on the device. When the local model cannot complete a step, the harness selects the relevant context, flags potentially sensitive information, shows what would leave the machine and asks for approval; the cloud model returns text guidance and gets no access to local files or tools.

Computer scored 65.1% on ParseBench-100 against 34.6% for Hermes

ParseBench-100 is a 100-task subset covering charts, layouts, tables, text and formatting. Computer scored 65.1% on it, Hermes 34.6% and Pi 13.9%. Perplexity says it plans to open-source the internal benchmark it used for these comparisons.

In one demo, Portable Computer reviewed a folder of tax documents on a DGX Spark; in a second, the local agent analysed a local CSV file and posted its findings to Slack. Connectors ship for Google Drive, Gmail, Slack and GitHub, and web searches and connector calls leave the device while model inference and private-document processing stay local.

Despite the name, Portable Computer is not Personal Computer, Perplexity's Mac and Windows application for working with local files and native apps. Kupp says the product runs the model and harness locally but does not do computer use at this point, and that the appeal is not having to fiddle around with inference.

Windows support due in September

Windows support is scheduled to follow in September, and Portable Computer is not available on Mac hardware. Kupp says Perplexity is for now very focused on Nvidia across DGX and RTX, while other hardware is under consideration. No date has been given for the arrival of Nemotron 3.5 Lightning or for the release of the internal benchmark.