What’s new here

Moondream ships two model sizes: 2B and 0.5B parameters. The 0.5B variant is framed as a distillation target for edge devices, enabling deployment on resource-constrained hardware while maintaining image understanding capabilities.

What it does

Moondream takes an image and answers questions about it. It handles visual question answering, image captioning, and object detection. You ask “What is the girl doing?” and it answers in plain text.

Two weights are available: the 2B model for general-purpose image understanding, and the 0.5B model for constrained hardware where you need to trade some capability for memory and speed.

The project points to its own documentation site at moondream.ai for the current quickstart, and lists Modal as a cloud runtime with a linked example.

Who it’s for

Teams or solo developers who need image understanding in a product but prefer to run inference locally rather than call a hosted API. Embedded systems work and local pipelines are a natural fit, given the 0.5B model’s small footprint.

Try it

The README defers to the hosted docs for setup instructions. Start at moondream.ai/c/docs/quickstart for the current install steps. A Modal-based cloud example is also available at the linked moondream-examples repository.

How mature is it

10,100 stars, 798 forks, 24 contributors. The repo has no formal releases tagged in GitHub. The last recorded push was April 2026, with zero commits in the 90 days before the snapshot date. There are 168 open issues and 70 open pull requests. License is Apache-2.0.