Moondream is a vision language model that reads images and answers questions about them. You can ask it to caption a picture, point to or segment objects, detect things described in plain words, or answer a direct question about what it sees. The open models are available to download and run yourself, and the same models are offered as a hosted API.
The product has four parts that work together. The open models are free for commercial use, with 2B, 1B and 0.5B checkpoints on Hugging Face. Lens is a fine-tuning platform where you can adapt a model to your own images, either self-serve through an API or with a white-glove option where the team handles labelling and evaluation. Photon is an inference engine built to run the models quickly on hardware from Jetson edge devices to datacenter GPUs. Moondream Cloud is a hosted inference service with an OpenAI-compatible API.
The product is aimed at teams building visual AI into their own software, such as warehouse, manufacturing, logistics, advertising and camera-based systems. You can try the models in a playground before writing any code, then install the Python package and run the model locally or call the cloud API.
On pricing, the models and Photon are free to use. Cloud inference is charged per token, with Moondream 3.1 listed at $0.30 per million input tokens and $1.00 per million output tokens for real-time requests, and lower batch rates. Lens training is $0.60 per million training tokens. Paid workspace plans start at $0 for an individual developer, and a Team plan is listed at $350 per month or $300 per month billed annually. Paid plans add seats, shared resources and support. They do not change token pricing.
The company is M87 Labs, based in the USA.
