Stream Real-Time Video to a Vision Language Model

Basic idea

Live VLM WebUI is a browser interface for real-time vision language model (VLM) interaction and benchmarking. It streams video from a webcam to an OpenAI-compatible VLM backend and displays live analysis alongside inference latency and system utilization.

The interface supports local backends such as Ollama, vLLM, SGLang, and NVIDIA NIM, as well as compatible cloud APIs. This playbook uses Ollama for a straightforward local setup.

What you'll accomplish

You'll create a real-time video analysis environment on your hardware platform that can:

• Stream webcam video to a locally served VLM
• Apply preset or custom prompts to incoming frames
• Display model responses, inference latency, and system utilization
• Compare supported vision models through one browser interface

What to know before starting

Required:

• Basic Linux command-line experience
• Familiarity with browser permissions and local network addresses

Optional:

• Familiarity with REST APIs
• Basic knowledge of vision language models

Supported hardware platforms

Use the matrix below to confirm your hardware platform, recommended default local settings, and whether multi-node applies.

Prerequisites

Hardware requirements

• Supported hardware platform — see the matrix above
• A webcam available to the browser
• Sufficient available storage for Python packages and the selected VLM

Software requirements

• Python 3.10 or later: python3 --version
• Network access to download packages and models
• A Chromium-based browser or Firefox with access to https://<HARDWARE_IP>:8090
• Direct network connectivity between the browser and hardware platform for WebRTC

Find model recipes

Browse models supported by each backend in the Live VLM WebUI list of vision-language models.

Use the Instructions tab for the base Ollama workflow. Model memory and latency requirements vary, so confirm that another model fits available memory before downloading it.

Ancillary files

No local ancillary files are required. The application and supporting documentation are available in the Live VLM WebUI repository.

Time & risk

• Estimated time: 30 MIN (longer on first run due to model download)
• Risk level: Low
  • Port 8090 must be reachable from the browser
  • The self-signed HTTPS certificate requires a one-time browser exception
  • Model downloads consume storage and network bandwidth
• Rollback: Stop the application, then optionally uninstall it with pipx uninstall live-vlm-webui. This does not remove Ollama or downloaded models.
• Last Updated: 10/22/2026
  • Added 64 GB DGX Spark to the supported hardware matrix
  • Changed the validated default vision model to gemma3:4b and added troubleshooting for incompatible Ollama model architectures