- Cloudflare-hosted
- Vision
Moondream 3 is a fast, efficient 9B mixture-of-experts vision language model (2B active parameters) that delivers frontier-level visual reasoning for tasks like object detection, pointing, OCR, and structured output.
| Model Info | |
|---|---|
| Vision | Yes |
| Unit Pricing | $0.30 per M input tokens, $1.00 per M output tokens |
Parameters
task
stringdefault: queryenum: query, caption, point, detectWhich Moondream skill to run.image
stringInput image as a public HTTPS URL or base64 data URI. Optional for `query`; required for `caption`, `point`, and `detect`.question
stringdefault: What's in this image?Question for the `query` task.caption_length
stringdefault: normalenum: short, normal, longCaption length for the `caption` task.target
stringdefault: personObject phrase to locate for `point` and `detect` tasks (e.g. 'person wearing a red shirt').reasoning
booleandefault: trueEnable reasoning trace for the `query` task.temperature
numberdefault: 0.2minimum: 0maximum: 2Sampling temperature.top_p
numberdefault: 0.9minimum: 0maximum: 1Top-p (nucleus) sampling.max_tokens
integerdefault: 8192minimum: 1maximum: 28672Max tokens to generate for `query` and `caption`.max_objects
integerdefault: 150minimum: 1maximum: 500Max objects to return for `point` and `detect`.stream
booleandefault: trueReturn incremental tokens for `query` and `caption`. `point` and `detect` do not support streaming.finish_reason
stringReason the generation finished.▶metrics{}
objectanswer
stringAnswer text for the `query` task. Null for other tasks.caption
stringCaption text for the `caption` task. Null for other tasks.▶points[]
arrayLocated points for the `point` task. Null for other tasks.▶objects[]
arrayDetected bounding boxes for the `detect` task. Null for other tasks.▶reasoning{}
objectReasoning trace for the `query` task when reasoning=true. Null otherwise.