AI Potluck
Back to Gap Map Model components / Multimodal models

Moondream

M87 Labs
open weights / Overall score: 2.6

Moondream is M87 Labs' small vision-language model line, built to run quickly on local hardware. Moondream 3.1, a 9B mixture-of-experts model with 2B active parameters, answers questions, writes captions, and returns points and bounding boxes as structured output, with an optional reasoning mode. The earlier Moondream 2 is a 2B model.

Scored on Moondream 3.1, whose custom license forbids reselling general-purpose hosted Moondream inference. Moondream 2 is still published under Apache 2.0.

Openness

3 high confidence
3.0
weights
open(model.safetensors on the Hub, ungated)
data
described(the changelog names RL fine-tuning across 55 vision-language tasks
code
partial(inference code only
license
Moondream-Model-License-1.0(commercial use, modification and redistribution allowed

Moondream 3.1's weights may be used commercially, fine-tuned and redistributed, but its license forbids offering Moondream itself as a general-purpose hosted service. M87 Labs publishes neither the training data nor the training code. Moondream-Model-License-1.0 allows commercial use only within a bound, so the license tier is use_bounded.

Adoption

4 high confidence
4.0

Hugging Face downloads over the trailing 30 days, summed across the Moondream 2 and Moondream 3 checkpoints under the vikhyatk and moondream accounts. Nearly all of it is Moondream 2, and the hosted Moondream Cloud is not counted.

Capability

2 high confidence
2.0

Moondream reads documents and charts and returns object points and boxes quickly on modest hardware. It works on one image at a time with no video input, which puts it a step below MiniCPM-V.

Verified 2026-09-27