Perplexity has launched Portable Computer, a new version of its AI agent platform designed to run primarily on local hardware rather than relying on cloud infrastructure. Developed in collaboration with NVIDIA, the system allows users to analyze files, work with private codebases, and execute complex AI workflows while keeping sensitive information on their own machines. The product is initially available to Perplexity Pro and Max subscribers using NVIDIA DGX Spark systems, with the first release supporting Linux.
Bringing AI Workflows On Device
Portable Computer is designed for organizations and professionals handling information that may be unsuitable for continuous transmission to external cloud services, including confidential documents, proprietary code, and internal business data. The platform can search local files, summarize documents, analyze datasets, inspect code, and carry out multi-step tasks directly on the device without consuming Perplexity usage credits for work completed locally. Perplexity said the system reflects growing improvements in local AI models and computing hardware that are making more sophisticated workloads practical outside centralized data centers.
The platform currently runs on NVIDIA DGX Spark using either Qwen 3.8 27B or PPLX 27B, a version of the Qwen model post-trained by Perplexity for its local agent environment. NVIDIA Nemotron 3.5 Lightning, a 30-billion-parameter open model optimized for specialized task execution, is expected to be added to the model selection later. Core components including the orchestrator, planner, tool router, scheduler, task queue, and local search index also operate directly on the user's machine.
Cloud Escalation for Advanced Tasks
Although Portable Computer is built around local execution, Perplexity has retained the ability to use cloud resources when a task requires capabilities unavailable on the device. The local system can escalate work to Perplexity's cloud infrastructure for current web information, browser access, connected applications, deeper research, or access to more than 15 advanced AI models. When information needs to leave the device for cloud processing, the software asks the user for permission before transferring the relevant content.
This hybrid architecture is intended to let users protect confidential information while selectively drawing on external resources when necessary. For example, sensitive details from a private financial document could remain on the machine while cloud-based research retrieves current market comparisons or recent transactions. Perplexity said its local model is trained to complete as much of a task on the device as possible before determining whether external capabilities are required.
Connectors, Search, and Dictation
Portable Computer also supports integrations with services including Google Drive, Gmail, Slack, and GitHub, allowing users to combine local information with connected workplace applications. Perplexity says the system can perform workflows such as reviewing GitHub issues alongside reports received through Gmail before preparing information that can later be shared through Slack. The platform can also access Perplexity Search, wide research, or deep research when a task requires information beyond what is stored locally.
Voice interaction is handled locally through NVIDIA's Nemotron 3.5 automatic speech recognition model, enabling users to dictate instructions without sending audio recordings to the cloud. Transcription and subsequent actions involving local files remain on the device, which could be particularly relevant when working with confidential client information or proprietary source code. Perplexity also said code and tool execution are isolated within sandboxed environments with controlled access to files and connected applications.
NVIDIA DGX Spark as the Initial Platform
Portable Computer is launching first on NVIDIA DGX Spark, a compact AI computing system built around NVIDIA's Grace Blackwell GB10 platform. DGX Spark combines a 20-core Arm processor with an NVIDIA GPU and 128 GB of unified memory, providing the computing capacity required to run relatively large AI models locally. Perplexity said users can install Portable Computer through a one-click setup available through its application.
The initial release is limited to Linux, although Windows support is planned, and Perplexity intends to expand availability to PCs equipped with NVIDIA RTX GPUs. The companies are positioning the launch as an early step toward broader adoption of locally operated AI agents as model performance improves and consumer and professional hardware becomes more capable. Running more workloads directly on personal or enterprise machines could also reduce dependence on usage-based cloud computing while giving users greater control over sensitive information.
Perplexity's Portable Computer introduces a local-first approach to AI agents that combines on-device execution with optional access to cloud-based research and reasoning. By working with NVIDIA and launching initially on DGX Spark, Perplexity is targeting users who need advanced AI capabilities while maintaining tighter control over confidential files, code, audio, and business data. With Windows and RTX support planned, the company is signaling that locally operated AI workflows could become a larger part of professional computing as hardware and open models continue to advance.