Molmo

Molmo is the Allen Institute for AI's open family of multimodal models, released in September 2024 as proof that open vision-language models could stand with proprietary ones. The family spans Molmo-72B built on Qwen2-72B, two 7B variants (one on Qwen2-7B, one on Ai2's own OLMo-7B, the most open of the set), and MolmoE-1B, a mixture-of-experts model with 1B active parameters that Ai2 said nearly matched GPT-4V. The 72B flagship topped 11 academic benchmarks and ranked just behind GPT-4o in user preference at release.

Molmo's signature innovation is pointing data: it is trained to point at objects in images, making it unusually strong at visual grounding, a capability Google DeepMind robotics researchers singled out as a step change for open VLMs. As with OLMo, Ai2 released weights, training data, fine-tuning data and source code, all under permissive Apache 2.0 licensing. openWeights is true. For developers needing a capable open vision-language model without proprietary API dependence, Molmo remains one of the strongest options.