Moondream 3 Preview
Neural Network
Moondream 3 Preview is a vision-language model by Moondream for image Q&A, descriptions, and OCR.
Specifications
- Input
- text, images, documents
- Output
- text
- Context size
- 32 000 tokens
- Max answer length
- 4 096 tokens
- Reasoning effort
- low, medium, high
- Capabilities
- reasoning
Prompt cost
(per 1M tokens)
Answer cost
(per 1M tokens)
How it works Moondream 3 Preview?
Moondream 3 Preview is a compact vision-language model from the Moondream team: 9 billion parameters total and 2 billion active due to the mixture-of-experts architecture, allowing the model to answer questions about images, write descriptions, recognize text in pictures, and perform visual reasoning without requiring the resources of large multimodal systems. In BotHub, it is available from Russia without a VPN or foreign card; payment is made in rubles using Russian cards based on actual usage. All 250+ models are gathered in one window: you can compare the Moondream response with another model and switch via a unified OpenAI-compatible API without rewriting the integration. Data is transmitted with AES-GCM encryption. It is useful if you need to parse a scan or photograph of a document and extract text, automatically caption product catalogs, build user image verification based on specific questions, collect descriptions for interface accessibility, or prototype an agent that reads charts and screenshots.Frequently asked questions about Moondream 3 Preview
You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.
moondream3-preview is a vision-language model from Moondream: it accepts text, images, and documents as input and responds with text. It is suitable for describing image content, finding objects in photos, extracting data from scans or tables, and answering questions about document pages. Context is 32,000 tokens.
Up to 4096 tokens per response, which is about a few pages. This volume is sufficient for image captions, extracted fields, descriptions, and short reports. If you need a longer result, break the task into parts and collect answers sequentially, passing the previous fragment into the context.
Yes, reasoning before answering is confirmed in our data: the model can break down a task into steps and only then provide a result. This helps with image questions where you need to compare several details, count objects, or link numbers from a document to the question itself.
In our data, function calling and structured JSON output are not noted for this model, so do not count on them in advance — test with your request. You can describe the required schema directly in the prompt, and if the format is critical, switch to another model via the same BotHub API.
Images and documents — yes, this is its main scenario: upload a photo, screenshot, or scan and ask about the content. Audio and video input are not noted in our data. The model responds with text; image or sound generation is not claimed.
We have no data on prompt and response languages, so we cannot promise the quality of Russian. The fastest way is to send your typical request along with an image and see the result. Other models are available nearby in the same BotHub window, so you can compare answers in a minute.