Moondream
Moondream is a small vision model with five typed skills. OpenTryOn uses the first-party cloud API and, for 3.1, the official Photon runtime.
| CLI id | Service | Upstream id | Where it runs |
|---|---|---|---|
moondream-3.1 | understand, detect | moondream3.1-9B-A2B | Cloud, MOONDREAM_API_KEY |
moondream-3-preview | understand, detect | moondream3-preview | Cloud, same key |
moondream-3.1-local | understand, detect | moondream3.1-9B-A2B | Photon, pip install moondream |
Auth header on the cloud API is X-Moondream-Auth. Optional MOONDREAM_BASE_URL defaults to https://api.moondream.ai/v1.
Understand
--skill query (default) needs --prompt. --skill caption uses --length short|normal|long.
opentryon understand --model moondream-3.1 --image look.jpg --prompt "What fabric is the jacket?"
opentryon understand --model moondream-3.1 --image look.jpg --skill caption --length short
Detect
Coordinates are normalized from 0 to 1, origin at the top-left.
opentryon detect --model moondream-3.1 --image look.jpg --object dress
opentryon detect --model moondream-3.1 --image look.jpg --object button --skill point
opentryon detect --model moondream-3.1 --image look.jpg --object jacket --skill segment
detect writes JSON boxes {label, score, bbox_xyxy}. point writes centroids. segment writes an SVG path. This is open-vocabulary grounding, not the garment-parser masks tracked in the new services backlog.
Local
Photon runs on NVIDIA Ampere or newer, and on Apple Silicon. Base weights do not need an API key.
pip install moondream
opentryon understand --model moondream-3.1-local --image look.jpg --prompt "Describe the outfit."
License: Moondream License. See the 3.1 model card.