Molmo

Molmo demonstrated that fully open training data could rival closed vision-language models: Ai2 released the family alongside PixMo, a collection of eight datasets for dense captioning, instruction tuning, grounding, and counting built without distilling from proprietary VLMs. Models span MolmoE-1B, a 7B-total mixture-of-experts with 1B active parameters, the dense Molmo-7B-O and Molmo-7B-D, and Molmo-72B, which Ai2 reported comparing favorably with GPT-4o on academic benchmarks and human preference at its 2024 debut. A distinctive capability is pointing: the models can output image coordinates, enabling counting and UI-grounding tasks most VLMs cannot do. The GitHub repository holds the Apache-2.0 modeling, training, and evaluation code released in December 2024, with checkpoints on Hugging Face and the PixMo datasets published separately. Everything needed to reproduce or extend the models is public, which remains rare among VLM releases. Serving the 7B variants needs a mid-range GPU, while the 72B model calls for multi-GPU hardware, and the 7B models are the most widely used in practice.