Favicon of Moondream

Moondream

An open-weight vision model for image questions, captions and object detection. Run it locally or use hosted inference and fine-tuning.

Screenshot of Moondream website

Moondream is a vision model for developers building software that needs to understand images. It can answer questions about a picture, write captions, locate objects, identify points and segment regions. The open-weight models can run on your own hardware, including in an air-gapped environment. The repository code is licensed under Apache 2.0; check each model checkpoint’s own terms for use.

There are models for different hardware needs. Moondream 3.1 uses a mixture-of-experts design for visual reasoning and longer queries. Moondream 2 is a smaller choice that runs on CPUs, GPUs and edge devices, while Moondream 2 0.5B is aimed at more constrained hardware and fine-tuning. Photon, Moondream’s inference engine, runs models locally on Mac, Windows and CUDA hardware, including NVIDIA Jetson devices. You can also self-host inference on a server.

Moondream has cloud services alongside the local models. Moondream Cloud provides hosted inference through an OpenAI-compatible API, so requests sent there are processed off your machine. Lens is a hosted fine-tuning service for adapting a model to images from a specific use case; it doesn’t require dataset uploads, and you keep the resulting weights. A tuned model can run through Moondream Cloud or locally with Photon. That gives teams working with specialized images, such as warehouse or manufacturing footage, a path from an open model to a model trained for their own visual tasks.

Similar to Moondream