Molmo 72B

Molmo 72B is the flagship of the Molmo family of open vision-language models, released by the Allen Institute for AI on September 24, 2024. Where most large multimodal models train on billions of indiscriminately scraped image-text pairs, Molmo was trained on only about 600,000 carefully curated images with human-annotated captions built without the use of other VLMs. Ai2 reported that the 72B model outperformed OpenAI's GPT-4o on tests of image, chart, and document understanding, and MIT Technology Review covered the release as evidence that open-source AI had caught up with closed labs.

The honest picture: Ai2 acknowledged that certain elements of the most powerful model remain shielded from view, so it is not as fully open as the 7B variants. It is also a 2024-era model, superseded by newer open VLMs on several benchmarks. But its data-curation philosophy — quality over scale — influenced a generation of open multimodal work, and the weights remain freely available on Hugging Face.