smolvlm-realtime-webcam

Why we included this project

This demo shows how to run a vision-language model against a live webcam feed on an ordinary laptop. SmolVLM-500M is small enough for that, and the project connects it to llama.cpp's server, then serves a minimal HTML page that streams camera frames to the model and displays detections in real time. Setup is a single server command plus opening one file, and because the instruction prompt is editable, you can ask the model to describe the scene or return JSON rather than just label objects. That combination makes it useful as a quick test of whether a small VLM fits your object-detection or instruction-following use case, and a compact reference for llama.cpp's multimodal API if you are exploring local, offline inference.