Perplexity Portable Computer Runs a Local AI Agent

What Perplexity Portable Computer keeps on your machine

Portable Computer keeps private data and the agent's working parts on your machine, and on-device work does not consume credits. Launched on August 25, 2026, it is a version of Perplexity Computer that runs entirely on a local machine. Perplexity built it with NVIDIA for the DGX Spark, NVIDIA's desktop-sized AI computer.

The post opens by noting that some of the most valuable work for AI agents involves data people would rather keep on their own machines, such as private codebases or confidential material. On the device, users can analyze data, synthesize files and run complex workflows.

The model is not the only part that stays local. The orchestrator, planner, tool router, scheduler, durable task queue and local search index all run on device as well. Perplexity adds that owning the compute makes high-volume tasks practical, because work handled by the local models has no per-credit charge.

Which models run locally on the DGX Spark?

Portable Computer runs Qwen 3.8 27B or PPLX 27B, a post-trained version of the Qwen model, on the DGX Spark. NVIDIA Nemotron 3.5 Lightning, a 30B open model, is listed as coming soon to the model picker. Dictation also runs locally, using the NVIDIA Nemotron 3.5 ASR Model.

The post says the Qwen model is post-trained to complete as much of each task locally as possible and to escalate to the cloud when the task needs it. It reads local files, searches across documents and code, takes actions on the device and keeps jobs running. The table below sets out each model and its status.

Models named in the launch post

Model	Status in the post	Where it appears
Qwen 3.8 27B	Runs on the DGX Spark	Local-first AI stack
PPLX 27B	Runs on the DGX Spark; a post-trained version of the Qwen model	Local-first AI stack
NVIDIA Nemotron 3.5 Lightning (30B open model)	Coming soon to the model picker	Local-first AI stack
NVIDIA Nemotron 3.5 ASR Model	Runs dictation locally	Search and dictation

The post attributes the local-first post-training to the Qwen model, so it does not single out PPLX 27B. It gives no date for Nemotron 3.5 Lightning. With dictation, the transcription and the actions on files stay on the machine, and the audio does not touch the cloud.

How cloud escalation works and when it asks first

When a task needs to send content from the device to a cloud service, Portable Computer asks the user for permission before moving forward. The post says the tool handles the majority of tasks locally, while some jobs benefit from the web or from frontier reasoning in the cloud. Its example is a term sheet: a user might keep the confidential details local but escalate to the cloud for current market comps or recent precedent deals.

The local orchestrator can escalate a task to the cloud for current information, browser use, connected apps, or one of more than 15 frontier models for advanced reasoning. For outside information, Portable Computer can also run Perplexity search or wide or deep research. The user gets cited findings from the web while everything else still runs locally.

App connectors cover Google Drive, Gmail, Slack and GitHub. In Perplexity's example, a user can triage new GitHub issues against overnight bug reports in Gmail on device, then send a Slack message about the top three issues with suggested owners by morning. Code and tool execution run in isolated sandbox environments with controlled access to files and connected apps, and security matches the cloud version of Perplexity Computer.

Hardware, plan and operating system you need

Portable Computer is available to Pro and Max subscribers on the NVIDIA DGX Spark, and the first release is on Linux. RTX GPU PCs and Windows support are both listed as coming soon.

The DGX Spark is built around the Grace Blackwell GB10 platform, which pairs a 20-core Arm CPU with an NVIDIA GPU and 128 GB of unified memory. Installation is a one-click setup in the Perplexity app. The table sets out each requirement as the post states it.

Requirements as published

Item	Requirement	Where it appears
Machine	NVIDIA DGX Spark; RTX GPU PCs coming soon	Introduction and closing section
DGX Spark specification	Grace Blackwell GB10, 20-core Arm CPU and NVIDIA GPU, 128 GB unified memory	Closing section
Plan	Pro or Max subscription	Closing section
Operating system	Linux in the first release; Windows coming soon	Closing section
Setup	One-click setup via the Perplexity app	Closing section

Pro and Max are the only plans the post names. It gives no date for Windows or for RTX GPU PCs.

What the launch post leaves unsaid

The August 25 post gives no price, no VRAM figure and no dates for Windows or RTX GPU PCs. The only hardware numbers it gives are for the DGX Spark: a 20-core Arm CPU and 128 GB of unified memory. Anyone weighing a purchase has to look to other sources for the rest.

The post links a separate research blog, A Local-First Agent for Private and Cost-Effective Knowledge Work, which Perplexity says details the development behind Portable Computer and its outperformance across benchmarks for accuracy, speed and credit efficiency. The launch post itself gives no figures for those claims.