Run Perplexity Locally on NVIDIA DGX Spark | Private AI with Local LLMs
Published 03/09/2026
Description
Run Perplexity locally on an NVIDIA DGX Spark and combine private, local AI inference with the flexibility of online AI models. Contents: 0:00 - Introduction 0:28 - Description 1:30 - Installing Perplexity 3:13 - First Run 4:38 - Testing 7:40 - Startup and Memory 9:31 - Local vs Online 11:33 - Config Overview 13:50 - Conclusion In this video, I install Perplexity Computer on an NVIDIA DGX Spark, download a local 27-billion-parameter model and local inference engine, and test how Perplexity can analyse documents locally without requiring sensitive data to leave the network. Perplexity is widely known for its online AI models and web-based research experience. With local inference on compatible NVIDIA hardware, it can also provide a hybrid AI workflow: run appropriate tasks locally for privacy and cost control, then optionally escalate more demanding work to online models when needed. I use a collection of simulated confidential project documents—including requirements, architectural decisions, security checks, risk registers, cost models, stakeholder communications, incidents, and contracts—to demonstrate a local executive project review workflow. The local model is asked to create a no-go recommendation, identify high-priority issues, find contradictions and unresolved assumptions, highlight security and network-exposure concerns, and recommend decisions for a steering committee. What you will learn How to install Perplexity Computer on NVIDIA DGX Spark How to add the Perplexity APT repository and install the application on Linux How to sign in to Perplexity Computer How to download a local model and local inference engine How local 27B-parameter AI models run on DGX Spark How to use local files and folders as AI input How to analyse potentially confidential data locally How to monitor GPU activity and unified memory usage How local inference differs from online Perplexity usage How local sessions and projects differ fro