Give your webcam a brain: local AI that reads, counts and reacts
A 1.3GB vision model on your laptop can read a parcel label, spot a fault on a display and tell you whether a car park is busy, in about four seconds, with nothing uploaded. Learn what small vision models are good and bad at with a real benchmark, fix their worst habit (counting), then set a watcher on your webcam or screen that saves a snapshot and alerts you when something you care about happens.
- A watcher for your webcam, screen or a folder: typed answers, rules in code, alerts, a log and snapshots
- A benchmark you can fill with your own images before you trust a model with a real job
- Ollama: Run AI on your PC. The models in this mission are 1.3GB to 4.6GB
- uv from docs.astral.sh/uv ↗
- A webcam for the live watcher (optional: everything else works with image files)
- 8GB of RAM for
qwen3.5:0.8b, 16GB for the bigger models
Meet the small vision models
Vision-language models answer questions about pictures. Pick a small one that runs well on your machine.
A vision-language model (VLM) takes an image and a question and answers in text: "what does this sign say?", "is there a parcel on the step?", "what temperature is on that display?". Ollama runs several, and
ollama showlists vision under Capabilities for the ones that can see.- qwen3.5:0.8b (1.3GB): the smallest Qwen 3.5. Fast enough for a laptop CPU, surprisingly good at reading text.
- gemma4:e2b (4.6GB download): Google's small Gemma 4, which also handles audio. The model from Run AI on your PC.
- qwen3.5:4b (3.3GB): slower on a CPU, more careful with details.
Terminalollama pull qwen3.5:0.8b ollama pull gemma4:e2b ollama pull qwen3.5:4b ollama show qwen3.5:0.8bGet the kitvision-kit.zip has four scripts, eight test questions and seven test images (four Wikimedia Commons photos with credits, three made for this mission). kit.md has the rule-writing guide and a camera etiquette checklist.
ollama show lists vision under Capabilities for the models you pulled.