Molmo

Molmo stands out for how much Ai2 shares. Many open-weight models release only the final weights, but Ai2 also publishes training datasets, code and recipes, and one Molmo 2 variant uses a fully open language backbone so every component can be examined. Its pointing ability, where the model marks the exact spot it is talking about, makes answers easier to check.

• Molmo 2 handles single images, multiple images and video, including tracking objects across frames.
• Pointing and counting let the model show where things are instead of only describing them.
• The 4B and 8B models are sized to run on a strong workstation for research and prototyping.
• Ai2 released new open datasets for dense video captions, long-form video questions and tracking.
• A free playground lets anyone try the models in a browser without setup.

Molmo is best for researchers, students and developers building vision features who value transparency and want to fine tune or audit their models. A practical tip: start in the Ai2 playground with your own images or short clips to see how pointing performs on your use case, then download the matching variant from the Molmo page when you are ready to run it locally.