Molmo: Free And Open Source Alternatives To Tech Giants’ AI

The Allen Institute for AI (Ai2) is a non-profit research organization based in Seattle. They have announced Molmo, a family of multimodal AI models. These models are designed to compete with offerings from tech giants like OpenAI, Google, and Anthropic, as reported by MIT Technology Review.

The term “multimodal” refers to the capability of handling various types of data, including text, images, audio, video, and even sensory information.

With the release of Molmo, Ai2 aims to provide an alternative that is not only free but also truly open source, as reported by TechCrunch (TC).

Among the Molmo models, the largest one has 72 billion parameters. Ai2 claims that this model can outperform OpenAI’s GPT-4o. GPT-4o is estimated to have over a trillion parameters. This performance comparison is based on tests that assess comprehension of images, charts, and documents, as reported by MIT Technology Review.

Ai2 attributes the success of Molmo to its innovative training methods, which prioritize high-quality data rather than relying on vast quantities.

Indeed, Ai2’s focus is on a more selective dataset, composed of just 600,000 carefully chosen images, minimizing inaccuracies commonly found in traditional models that utilize large datasets harvested from the internet, as noted by TC.

Ani Kembhavi, a senior director of research at Ai2, noted that this approach helps to reduce errors and confusion in AI outputs, as reported by MIT.

For those interested in exploring the model, a public demo is available that also can be accessed via mobile devices.

Importantly, Molmo is entirely free and open source, capable of running locally without the need for extensive computing resources. The intention behind this release is to empower developers and creators to build AI-driven applications and services without the necessity of obtaining permissions from or paying large tech corporations, as noted by TC.

“We’re targeting, researchers, developers, app developers, people who don’t know how to deal with these [large] models. A key principle in targeting such a wide range of audience is the key principle that we’ve been pushing for a while, which is: make it more accessible,” said Ali Farhadi, the CEO of Ai2.

He emphasized the organization’s commitment to accessibility, stating, “We’re releasing every single thing that we’ve done. This includes data, cleaning, annotations, training, code, checkpoints, evaluation. We’re releasing everything about it that we have developed.”

WIRED notes that creating effective AI agents requires reliable multimodal models. Enhancing reasoning abilities, as demonstrated by OpenAI’s o1 model, may be key to achieving this. The next challenge is integrating such reasoning into multimodal models.

This comprehensive approach aims to democratize access to advanced AI technologies, allowing a broader audience to innovate in this rapidly evolving field.