Molmo2 4B
Molmo2 4B understands images, multiple images and video.
What the release provides
Its language backbone is Qwen3-4B-Instruct-2507, paired with a SigLIP vision encoder. The Apache 2.0 release supports visual grounding. The card distinguishes released datasets from other training artifacts planned for later publication.Molmo2 4B official model card
How to use it well
Practical guidance. Keep frame selection and image preprocessing fixed when comparing results. For counting or grounding, inspect the returned locations against the original image rather than accepting only a fluent description.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | allenai/Molmo2-4B |