The Allen Institute for AI on Wednesday introduced Molmo, a family of open multimodal AI models. Molmo can understand visual data from everyday objects and signs, the Allen Institute for AI (Ai2) said.
In a video, the nonprofit research institute shows a Molmo model understanding and responding to various images and objects. Users show Molmo an image such as a parking sign, ask it a question and the model can understand what the sign means.
Molmo models also can point to what they perceive. The models can point to UI elements on the screens for developers.
Ai2 said it plans to open up Molmo's language and vision training data, fine-tuning data, model weights and source code in the future. However, some model weights, inference code and demos are available starting today.
The models come in different sizes: Molmo 72B, Molmo 7B-D, Molmo 7B-O and Molmo E-1B. The Molmo-1B model is tiny and can fit on most devices, Ai2 said.
Open models
The introduction of Molmo highlights the small gap in popularity between open and closed models in the generative AI market.
"Open in the world of AI is getting off to a running start in a way that open in, say, the operating system world didn't," Futurum Group analyst David Nicholson said.
In the operating system market, it took years before open source systems like Android and Linux caught up to proprietary systems like Mac OS and Windows OS. In contrast, open source -- in which the vendor releases source code -- and near open source -- or just open -- has already caught up with closed source in the generative AI market. Open models from Meta and independent generative AI vendor Mistral have gained popularity.
For example, according to Ai2, its 72B Molmo model is on par with the OpenAI GPT 4o and Google Gemini 1.5 proprietary LLMs in terms of performance.
Typically, if a vendor is truly open, it compromises in performance, Nicholson said.
"Unless they completely made this up, it's remarkable that they are willing to publish all of the information about their models while delivering the kind of performance that they claim that they are," Nicholson added.
Visual data
Ai2's willingness to make its data open is also noteworthy, Gartner analyst Arun Chandrasekaran said.
"The more transparent companies are in this space -- particularly academic institutions like the Allen Institute of AI -- the better it is," Chandrasekaran said.
The more transparent companies are in this space -- particularly academic institutions like the Allen Institute of AI -- the better it is.
Arun ChandrasekaranAnalyst, Gartner
Ai2's focus on vision with the Molmo models able to point to and understand the outside world is the way for AI models to get better and smarter, Nicholson said.
"Training these systems to understand what they 'see' is really, really critical to making them smarter," he said.
Ai2 is also focusing on the ability of the models to act as autonomous agents. In another video presentation, a Molmo model made a food order and scheduled it for pickup.
"If what these folks are saying is true, then this is a completely open set of tools that people can use to build their own agentic AI, not just generative LLMs," Nicholson said.
Some challenges
One hurdle Ai2 faces is the need to create an ecosystem around its models like what Meta and Mistral are building, Chandrasekaran said.
"It's one thing to make these models very accessible to the developers. But it's entirely another thing to build a strong community around it and really think about enterprise needs and how they can galvanize this ecosystem around enterprise needs," he said. "That's critical."
He added that the marketplace already has a lot of good models. Vendors must now think of how to build on top of the models so enterprises can deploy the models more efficiently, he said.
"They have to think more up the stack in terms of the platform tools that they can combine with these models that makes them useful for enterprises across a broad set of use cases," he added.
It's also unclear whether enterprises will embrace Ai2 and its models, Nicholson said.
"This is starting out as like a bit of a philanthropic effort, which is really interesting, and extremely important right now to offset what the commercialization of AI is involved with," he said. "It's unclear to see whether it will be embraced like Linux was embraced."
Esther Ajao is a TechTarget Editorial news writer and podcast host covering artificial intelligence software and systems.
