Ai2 releases Molmo 2 for video understanding

Background

Ai2's first Molmo models worked on still images. According to Ai2, that original release set a new state of the art on static-image benchmarks and pioneered pointing as a way for a model to answer ^[2]. Pointing means the model does not only write a sentence in reply; it also indicates a location in the picture, for example the pixel where an object sits. That makes an answer checkable, and it gives software a concrete coordinate to act on rather than a description that still has to be parsed.

What was released

On 11 December 2025 Ai2 released Molmo 2 ^[1]. The family comes in three sizes: an 8B model, a 4B model, and a 7B version built on Olmo ^[1]. Olmo is Ai2's own line of open language models, so the 7B variant pairs a Molmo vision component with an Olmo language backbone.

The weights are released under the Apache 2.0 licence ^[1], which permits use, modification and redistribution, including commercial use, provided the licence terms are kept. For teams that want to run or adapt a vision-language model themselves, the licence is often as important as the benchmark numbers: it determines whether the model can be built into a product or fine-tuned on private data without negotiating a separate agreement.

What is new

Molmo 2 extends the earlier image work to video. It adds video pointing, tracking and grounded question answering ^[1].

• Video pointing is the moving-image version of the original capability: the model localises something across frames instead of marking it once in a single picture.
• Tracking means following the same object through a sequence as it moves, is partly hidden, or changes appearance.
• Grounded question answering means the answer is tied to a specific region the model has identified, rather than being free-floating text about the scene as a whole.

Ai2 said the 8B model tracks objects in video better than Gemini 3 Pro ^[1]. That is Ai2's own comparison, and it concerns tracking specifically rather than overall model quality.

The codebase

The Molmo 2 codebase was open-sourced on 3 March 2026 ^[2]. It covers data preparation, pretraining, fine-tuning and evaluation ^[2], that is, the whole pipeline from assembling training data to measuring how well the finished model performs. Releasing a codebase alongside weights matters because weights alone let others use a model, while the pipeline lets them rebuild it, check it, or reproduce the reported results.

Why it matters

Molmo 2 sits at the intersection of two trends. One is the shift of vision-language models from still images to video, where an object has to be located and followed over time rather than found once. The other is open release: permissive weights plus a published training and evaluation pipeline make the model something others can inspect and extend, not only call through an interface. Pointing and tracking also suggest a different kind of use, in which a model's output is a region or a trajectory in the world rather than a paragraph of text.