Moondream

Moondream squeezes vision-language capability into models small enough for edge deployment, now led by Moondream 3.1, a 9B mixture-of-experts with only 2B active parameters and a 32K context, alongside the earlier Moondream 2B and a 0.5B variant for tightly constrained hardware. Four native skills return structured output: caption, query for visual question answering, detect for open-vocabulary object detection, and point for pixel coordinates, which also covers counting and reading text off images. The repository runs locally on CPUs or modest GPUs, and M87 Labs also operates a hosted cloud API with Python and Node.js SDKs plus plain REST endpoints documented at docs.moondream.ai. Licensing splits by generation: the code and the 2B and 0.5B weights are Apache-2.0, while Moondream 3.x weights use the Moondream Model License derived from Business Source License 1.1, which permits internal and most commercial use but bars reselling hosted access to the model itself. Roughly 9.9k GitHub stars and steady releases since 2024 have made it a default for image understanding inside resource-constrained products.