Moondream is an open-weight vision-language model with a small footprint, aimed at image understanding tasks an agent pipeline can call directly: querying an image in natural language, captioning, detecting objects, pointing at them, and segmenting. Detection returns normalised bounding-box coordinates, so results can be consumed as structured data rather than prose.
Several variants are published. The 2B model is the general-purpose workhorse for captioning, visual question answering and object detection, and the 0.5B model is a compact distillation target for edge devices where memory is constrained. A larger sparse mixture-of-experts variant carries 9B parameters with 2B active, and adds segmentation and long-context queries. Models can be run locally or through a hosted API, and a playground exercises them in the browser.
Two adjacent products surround the models: Photon, the inference engine used to run Moondream in production and which added speech recognition in the 2.1 release, and Lens, a hosted fine-tuning API supporting supervised fine-tuning and reinforcement learning. Lens is documented as improving accuracy from as few as 20 labelled images, with the resulting weights, training code and data staying with you and the tuned model runnable on the hosted cloud or locally on Photon.
Features
- Query and caption: answers natural-language questions about an image and generates captions of varying length
- Detect, point, segment: locates objects by description and returns normalised coordinates or masks
- Model sizes: a 2B general-purpose model, a 0.5B edge and fine-tuning base, and a 9B mixture-of-experts variant with 2B active parameters
- Runs anywhere: local execution on GPUs, CPUs and edge devices, or a hosted cloud API
- Photon: the inference engine behind production deployments, with speech recognition added in the 2.1 release
- Lens fine-tuning: a hosted API for supervised and reinforcement-learning fine-tuning from small labelled sets, with no dataset upload or infrastructure to manage
- Open weights: the open models are published for commercial use
