The Allen Institute for AI (Ai2) has released Molmo 2, a suite of open video-language models, together with the training data behind them — a combination that reflects the nonprofit’s continued commitment to open source and appeals to enterprises that want tighter control over the models they deploy. The release, announced in mid-December 2025 and detailed on Ai2’s blog, includes Molmo 2-4B and Molmo 2-8B, both built on Alibaba’s Qwen3 language model, plus Molmo 2-O-7B, a fully open version based on Ai2’s own Olmo language model.
Alongside the models, Ai2 published nine new datasets, including a long-form question-answering dataset for multi-image and video input and an open video pointing-and-tracking dataset.
What Molmo 2 can do
According to Ai2, the Olmo-based variant is transparent end to end: because users have access to both the vision-language model and its underlying LLM, they can study and customize the full stack — a level of openness rare among multimodal models. New capabilities include multi-image understanding and support for video of arbitrary length.
The models’ distinguishing feature is grounding. Ranjay Krishna, director of perceptual reasoning and interaction research at Ai2, explains that Molmo 2 does not just answer questions about images or video — it points to the specific pixels, or the moment in time, where something happens. The models can also generate descriptive captions, track and count objects across frames, and detect rare or surprising events in extended video sequences. In published evaluations, Ai2 reports that the 4B and 8B models lead both proprietary and open-weight vision-language models on multi-domain object tracking. Molmo 2 is available on Hugging Face and in the Ai2 Playground.
Why full openness matters to enterprises
Industry analyst Bradley Shimmin of Futurum Group argues the release underscores the value of vendors who publish not just model weights but the associated training data. As enterprises push corporate data into models under data-sovereignty requirements — where data must comply with the laws of the country where it was generated — that transparency becomes a compliance asset, not just a philosophical stance.
Shimmin also sees significance in Ai2’s decision to keep the models small, at four to eight billion parameters. Trillion-parameter frontier models are not economically viable for every enterprise, and in his assessment the Molmo family shows that frontier scale is not required to extract value. Enterprises, he notes, increasingly recognize that training data matters more than parameter count, and many now demand transparency and accountability about the data a model is built on — a key argument for the open-source innovation model across the IT landscape.
Limitations and what to watch
Open releases of this kind face structural headwinds. As Shimmin cautions, in an industry that allocates capital based on estimated future value, a nonprofit like Ai2 can be left behind on adoption and financing despite technical merit. Benchmark leadership claims — including comparisons against proprietary systems — come from Ai2’s own evaluations and the accompanying technical report, and independent replication will take time. Prospective users should also note the licensing difference within the family: the Qwen3-based variants inherit upstream licensing considerations, while the Olmo-based model is the fully open option. Worth watching: enterprise uptake of small open multimodal models, and whether grounded video understanding — pointing and tracking — becomes a standard capability across the field. Related reading on this site: Nvidia’s Nemotron 3 open models and Ai2’s open coding agents family.
