Moondream: Moondream 3 Preview

Moondream: Moondream 3 Preview is a vision-language model from Moondream, listed on api14 under the main category of text. It is a mixture-of-experts model with 9B total parameters and 2B active parameters, which means only part of the network is engaged for each request. The "preview" label in the name indicates that this is an early version of the third Moondream generation rather than a final release. It sits in the text category on api14 even though it accepts images, because its output is always text.

The model accepts text and image inputs and returns text only. Its context window is 32,000 tokens, and the maximum output is 4,096 tokens. According to Moondream, it handles visual question answering, captioning, OCR, and visual reasoning over images. Supported request parameters are include_reasoning, max_completion_tokens, max_tokens, reasoning, temperature, and top_p. The presence of reasoning and include_reasoning suggests the model can expose intermediate reasoning steps when you ask for them, though the exact form of that output is not specified here.

Those capabilities point to a fairly specific set of jobs. Visual question answering fits anywhere a user needs to ask about the contents of a photo or screenshot and get a written answer. Captioning suits generating alt text, image descriptions for catalogs, or summaries of visual material. OCR covers pulling text out of images, such as scans, receipts, or photographed documents, and returning it as plain text. Visual reasoning goes a step further, for tasks where the answer depends on combining what is visible with a question, such as interpreting a chart or checking whether an image matches a claim. Because output is text only, the model is not a fit when you need generated images.

You call it through api14's OpenAI-compatible API using the model id "moondream/moondream3-preview". Any client that speaks the OpenAI chat completions format should work; you set the model field to that id and send your text prompt along with any images you want analyzed. Tune sampling with temperature and top_p, cap the response with max_tokens or max_completion_tokens, and enable reasoning with the reasoning or include_reasoning parameters if you want the model's reasoning surfaced.

Keep the 32,000 token context and 4,096 token output cap in mind when designing prompts: long documents plus several images can fill the window quickly, and very long answers will be truncated. If your task is pure text with no images, or you need audio or video input, or you need image output, a different model on api14 will fit better than this one.