Molmo 2

About

The page introduces Molmo 2, a new family of open multimodal models developed by Ai2, designed for state-of-the-art video understanding, pointing, and tracking. Building on the original Molmo's success in image understanding, Molmo 2 extends these capabilities to video and multi-image inputs. It offers three variants (8B, 4B, and an Olmo-backed 7B) optimized for different needs, demonstrating superior performance and efficiency compared to its predecessors and some proprietary systems on key benchmarks like video tracking, image/multi-image reasoning, and video grounding. Molmo 2 supports advanced features such as video pointing, multi-object tracking with persistent IDs, dense video captioning, anomaly detection, and subtitle-aware QA. The model's open and extensible architecture combines a vision encoder with a language model backbone (Qwen 3 or Olmo) and was trained on a meticulously curated, video-centric multimodal corpus of over 9 million examples, including nine new datasets for dense captioning, long-form QA, and grounded pointing/tracking. It is intended for research and educational use.