Molmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding

Molmo2 is a new open-source family of vision-language models that achieves state-of-the-art performance among open models, particularly excelling in video understanding and pixel-level grounding tasks. The research introduces 7 new video datasets and 2 multi-image datasets collected without using proprietary VLMs, along with an 8B parameter model that outperforms existing open-weight models and even some proprietary models on specific tasks.

• →Molmo2 represents the first state-of-the-art open-source video-language model with full data transparency and no reliance on proprietary model distillation.
• →The model introduces exceptional point-driven grounding capabilities across single images, multi-images, and videos - a capability lacking in most proprietary models.
• →Seven new video datasets and two multi-image datasets were created specifically for training, including detailed video captions and innovative video pointing datasets.
• →The 8B model significantly outperforms existing open-weight models like Qwen3-VL and surpasses proprietary models like Gemini 3 Pro on video pointing and tracking tasks.
• →Novel technical innovations include bi-directional attention on vision tokens and a new token-weight strategy that improves performance efficiency.