Perplexity Portable Computer Launches for NVIDIA DGX Spark

Perplexity has launched a new “Portable Computer” system aimed at running its AI computing engine on a user’s own hardware instead of relying entirely on cloud-hosted models. The debut is set for NVIDIA’s DGX Spark platform, bringing built-in local execution plus controlled, on-demand cloud steps for harder tasks.

Portable Computer brings a complete local AI workflow

Portable Computer is designed to package Perplexity’s computing experience directly onto a local machine, removing the need for users to assemble components such as inference engines, agent harness configuration, sandboxing, and system tool hookups. The system includes local models, the inference engine, application connectors for Gmail, Google Drive, Slack, and GitHub, PDF reading, voice dictation, and an isolated, encapsulated sandbox environment for safer command execution.

A key focus is that sensitive data remains on the PC. The system supports direct interaction with local files without sending that information out of the device, targeting the privacy and operational-cost concerns that come with cloud-first AI setups.

Dynamic routing: local first, cloud only when needed

Portable Computer uses a dynamic routing approach so users don’t need to manually switch between models or create complex rules. By default, work starts on the user’s local hardware. If the local model encounters a task step that exceeds its reasoning or reach, the system prompts the user for approval before sending only that specific step to a cloud “frontier” model. The cloud output is then integrated back into the local execution flow in real time.

Local by default: all work begins on the user’s hardware
Targeted escalation: the system asks the user before sending only the required step to a cloud model
Seamless return: cloud results are merged back into the local run immediately
Cost optimization: local tasks don’t consume subscription tokens; only the cloud-used queries do

Launch models, NVIDIA hardware targets, and roadmap

At launch, Portable Computer supports open models including Qwen 3.8 27B and Qwen 3.6 35B, which Perplexity is said to be post-trained to improve agent task precision and reliability.

Upcoming additions include NVIDIA Nemotron 3.5 Lightning. The platform also allows users to bring their own models and inference servers for specific needs.

Initial availability is aimed at users on Pro, Max, Enterprise Pro, and Enterprise Max subscriptions running on Linux. NVIDIA DGX Spark is the first officially certified hardware. NVIDIA RTX GPU support on Linux is planned “soon,” while Windows support is scheduled for September.

The rollout ties into NVIDIA’s broader push for local AI, including optimization of open models with NVFP4, multi-token prediction (MTP), and management tooling such as NVIDIA Sync.

What to watch next

With the system debuting on DGX Spark and expanding to RTX support on Linux, the next milestones are the arrival of NVIDIA RTX compatibility and the Windows release in September—both of which will determine how widely Portable Computer can be used for local, privacy-focused AI agent workflows.