Perplexity Launches Portable Computer Agent Optimized for Nvidia DGX

What happened

The page returned mostly HTML/CSS/JS. Let me look for the actual article text content more specifically. The page is heavily JS/CSS-laden. Let me try using curl with a different approach to get the article text, or look for Perplexity-specific content: I now have the concrete facts from the source. Let me write the article:

Perplexity Launches Portable Computer Agent Optimized for Nvidia DGX Spark

Perplexity has released Portable Computer, a local AI agent application built to run on Nvidia DGX Spark. The app lets users operate a personal AI agent entirely on-device, keeping workflows private and, when running locally, outside any token-limit accounting.

How it works

This piece covers what Portable Computer is, how it works on DGX Spark, which integrations it ships with, what hardware it currently supports, and what Perplexity has disclosed about upcoming additions. It is written for developers and productivity workers evaluating local AI agents for daily or long-running autonomous tasks.

Perplexity announced Portable Computer on Tuesday, August 25, at 6:00 a.m. PT, as part of Nvidia's rolling blog series celebrating local AI in August. The announcement appeared on the Nvidia blog alongside a disclosure that the agent's local capabilities have been specifically optimized for DGX Spark, Nvidia's portable personal AI supercomputer. The app is downloadable directly from the Perplexity app distribution channel and is currently described as available now for DGX Spark users.

The framing of the release positions Portable Computer as a polished, consumer-friendly entry point into local agentic AI. Rather than a developer-facing SDK or API, Perplexity is presenting this as a finished product aimed at productivity workers who want capable agent behavior without configuring infrastructure. The combination of a named hardware target and a ready-made download marks this as a tighter hardware-software pairing than most local AI launches.

Why it matters

Portable Computer introduces one-click local inference setup on Nvidia hardware, removing the model-download and runtime-configuration steps that typically precede running large language models on a local machine. When running locally on DGX Spark, the agent uses a specially post-trained Qwen 3.8 27B model. The app also supports switching between local and cloud models, so a user can direct routine tasks to the on-device model while routing heavier or more complex requests to cloud-hosted endpoints.

The agent arrives with prebuilt integrations for Google Drive, Gmail, Slack, and GitHub. DGX Spark is described as providing fast large language model inference and 24-by-7 always-on operation suited to long-running autonomous agent tasks. Local inference on DGX Spark does not count against token limits under Perplexity's current scheme, which has direct cost implications for users running high-frequency or continuous workflows.

Productivity workers who currently run agent workflows through cloud APIs and want to move sensitive data off third-party servers are the primary audience named by Perplexity. The Gmail, Slack, Google Drive, and GitHub integrations suggest the tool is pointed at knowledge workers whose day-to-day context lives in those platforms. Because the agent can run continuously on DGX Spark without incurring token charges, anyone automating repetitive background tasks stands to benefit from the always-on architecture.

Who is affected

Builders evaluating this should verify two things before committing to DGX Spark as a deployment target: whether the post-trained Qwen 3.8 27B model meets their specific task quality bar, and how the app handles credential storage for connected services like Gmail and Slack. The switch between local and cloud modes also introduces a routing decision that may matter for compliance-sensitive environments where data residency rules apply to some requests but not others.

Portable Computer is available now for DGX Spark through a direct download from the Perplexity app. The Perplexity blog has additional setup documentation. Current hardware support is limited to DGX Spark; support for GeForce RTX and RTX PRO GPUs is listed as coming soon. Perplexity has also disclosed that Windows support and compatibility with DGX Station are in progress but has not given dates for either.

To try it on DGX Spark today, a user needs to download Portable Computer from Perplexity's app, complete the one-click local inference setup, and connect whichever of the four supported services — Google Drive, Gmail, Slack, GitHub — apply to their workflow. The local model that runs at setup is the post-trained Qwen 3.8 27B variant. The cloud-model option remains available for tasks requiring more compute than the local configuration provides.

What to watch next

Two named additions are confirmed but undated. Perplexity is working on a fine-tuned Nemotron 3.5 Lightning variant described as targeting ultra-fast responses, which will slot in alongside or replace the Qwen 3.8 27B as a local model option. GeForce RTX and RTX PRO GPU support is also described as coming soon, which would extend the addressable hardware base well beyond DGX Spark to the installed base of Nvidia consumer and workstation GPUs.

The DGX Station compatibility disclosure is the other signal worth tracking. DGX Station occupies a different tier from the portable DGX Spark, and support for it would position Portable Computer as a fixed-workstation agent rather than only a portable one. Builders already running Nemotron models should watch for the Nemotron 3.5 Lightning release closely, since Perplexity explicitly named the model rather than describing it generically, suggesting the integration is further along than a roadmap placeholder.