Moondream: Moondream 3.1

Moondream: Moondream 3.1 is a vision-language model from Moondream, the same team behind the earlier Moondream releases. It is the successor to Moondream 3 Preview, and it uses a mixture-of-experts design with 9B total parameters and 2B active parameters. That combination means the model carries a large overall capacity while only activating a smaller slice of it per request, which is a common way to keep inference efficient without shrinking the model's knowledge. It sits in the text category on api14, but its defining trait is that it reads images as well as text.

On the input side, Moondream 3.1 accepts both text and images, and it produces text only. Its context window is 32,000 tokens, and the maximum output length is 4,096 tokens. The provider describes the model as covering visual question answering, captioning, and OCR, among other tasks. Those are the core vision-language behaviours: answering questions about what is in an image, generating a written description of an image, and extracting text that appears inside an image. Because the context window is shared between the image input and the text prompt, larger or more numerous images leave less room for instructions and conversation history.

The practical use cases follow directly from those capabilities. Visual question answering suits interactive tools where a user uploads a photo or screenshot and asks about its contents. Captioning fits pipelines that need alt text, media descriptions, or short summaries of image assets. OCR is useful for pulling text out of scans, screenshots, forms, or photographs of documents, and for feeding that text into later steps. Because the model also takes ordinary text, it can handle a mixed prompt where the instruction and the image arrive together, which keeps integration simple for document and image review workflows.

You call it through api14's OpenAI-compatible API using the model id "moondream/moondream3.1-9b-a2b". The supported request parameters are include_reasoning, max_completion_tokens, max_tokens, reasoning, temperature, and top_p. Temperature and top_p control sampling, while the token parameters cap the length of the generated response. The reasoning and include_reasoning parameters relate to exposing the model's reasoning behaviour, so check how your client handles them if you do not need that output. Since the interface is OpenAI-compatible, existing SDKs and chat-style request shapes generally work without changes beyond the model id and the image content in the message.

Keep the limits in mind when choosing this model. The 32,000 token context and 4,096 token output ceiling are the hard boundaries, so very long documents or long multi-turn image conversations may need chunking or trimming. The output is text only, so it will not generate or edit images. If your task is purely text and does not involve images, a text-only model may be a simpler fit; Moondream 3.1 is most valuable when the request actually contains an image to interpret.