The Allen Institute for AI released Molmo 2, a suite of open video language models, on Tuesday. The new additions alongside training data show the non-profit’s continued commitment to open source, a benefit to enterprises looking to better control their use of the model.
The new models include Molmo 2-4B and Molmo 2-8B both built on Alibaba’s Qwen3 language model. The release also includes Molmo 2-O-7B, a fully open variant based on Ai2’s Olmo language model.
Along with the models, the nonprofit released nine new data sets including long-form quality assurance data sets for multi-image and video inputs, as well as an open video pointing and tracking data set.
Molmo 2’s capabilities
The Olma variant, Molmo 2-O-7B, is a transparent model that users can study end-to-end, according to the company. Because users have access to the vision language model and its LLM, Olmo, they can fully customize the model, providing a level of transparency, Ai2 said.
For the Molmo 2 models, one of the new capabilities Ai2 added is the ability to understand multiple images image; the model supports all images and any video, regardless of length, the company said.
Users can ask the model a question about the images or videos and it can base its reasoning on patterns it recognizes in the video, said Ranjay Krishna, director of perceptual reasoning and interaction research at Ai2.
“By ‘ground,’ I mean it’s not just giving you an answer, it’s sort of pointing out in the pixels or in time when something happens,” Krishna said.
The models also can generate descriptive captions, track and count objects across frames and detect rare or surprising events in extended video sequences. Molmo 2 is available on Hugging Face and Ai2 Playground, a platform from the non-profit where users can play with its different tools and models.
Commitment to Open Source
The release demonstrates Ai2’s continued commitment to open source; it highlights the importance of having a vendor that not only releases models but also the data and weights associated with them, according to Bradley Shimmin, an analyst at The Futurum Group.
“They should be given some attention, especially as we start to see an increasing emphasis on bringing corporate data to models in a way that emphasizes sovereignty,” where data must adhere to the laws of the country in which it was generated, Shimmin said.
He noted Ai2’s decision to keep its model sizes small — using four or eight billion parameters — is crucial because not every enterprise can afford or even needs a one trillion parameter model that has to be fine-tuned.
“It’s just not financially viable,” he said. “Molmo is such an important family of models that you don’t need to have a frontier scale to achieve value.”
Enterprises are also waking up to the fact that, instead of the size of a model, what matters is the data that it is trained on, he said.
“A lot of companies are demanding from the model makers that sort of level of transparency and accountability for not just the model itself, but the data on which it’s built, to be able to give them that freedom to innovate,” Shimmin said. “It’s another reason why the open source model of innovation is so critical to … the entire IT landscape.”
While the new Molmo 2 models provide greater flexibility for fine-tuning and high-quality data for those who want to iterate on it, they also present some challenges — namely adoption and, along with it, funding.
“When an industry like ours moves money according to perceived future value, it’s easy for a company like this to get left behind or left out,” Shimmin said.
