It displays structured bounding boxes over images; selecting a box reveals its label, confidence, and short reasoning. Replay mode is the default, requires no API key, and uses synthetic answers labeled illustrative. Live requests can be routed through Vercel AI Gateway, Mistral API, or OpenRouter.
The implementation requests coordinates normalized to 0–1000 because the author says Mistral’s launch materials and model card do not document a box convention. The pull request therefore demonstrates an application-level interpretation, not an official coordinate schema or validated production behavior. It says integrations should validate coordinate order, origin, scale, and behavior near image boundaries. The demo’s tests and successful build checks are reported in the pull request; they do not establish model accuracy across images.