Molmo2: Open-Source Vision-Language Models with State-of-the-Art Video Grounding

We present Molmo2, a new family of fully open state-of-the-art vision-language models (VLMs) developed by the Allen Institute for AI and the University of Washington. While most strong models today are proprietary, Molmo2 provides the community with open weights, open data, and open code. Key Highlights of Molmo2: • Superior Grounding: Molmo2 excels at point-driven grounding and object tracking in single images, multi-image sets, and videos. • Fully Open Data: Unlike other models, Molmo2's training data (including 9 novel datasets) was constructed from scratch without distilling from proprietary models like Gemini or GPT. • Performance: The 8B model significantly outperforms existing open-weight models and even surpasses proprietary systems like Gemini 3 Pro on specific tasks like video pointing and tracking. • Model Variants: The family includes Molmo2-4B, Molmo2-8B (based on Qwen3), and Molmo2-O-7B (built on the fully-open OLMo architecture). • New Capabilities: It handles dense video captioning, complex counting, and spatio-temporal localization with high precision. Learn more about the training recipe involving bi-directional attention, efficient sequence packing, and message-tree encoding that allows these models to reach top-tier performance while remaining transparent. Try the Demo: playground.allenai.org Github Code: github.com/allenai/molmo2