Molmo 2's architecture consists of a vision encoder that processes images or video frames into visual tokens and a language model backbone (Qwen 3 or Olmo) that consumes those tokens alongside text. A lightweight connector interleaves visual tokens with timestamps, image indices, and text so the model can reason jointly over space, time, and language. All variants are trained in two stages: first focusing on pretraining for alignment and grounding through joint image captioning and image pointing, then supervised fine-tuning on a multimodal mixture integrating images, multi-image sets, videos, and pure text across categories including captions, image QA, video QA, pointing, tracking, and NLP.

These capabilities open up a range of applications: counting-by-pointing, multi-object tracking with persistent IDs that follow objects across occlusions and re-entries, dense video captioning with highly descriptive and searchable narratives of long clips, anomaly detection that flags rare or surprising events, artifact detection that points to flaws in generative video such as inconsistent lighting or broken object geometry, and subtitle-aware QA that combines visual evidence with in-video subtitles. For prompts like 'Point out every instance where the person in the striped shirt flexes their muscles,' Molmo 2 analyzes the entire clip, emits coordinates and timestamps for each event, and maintains stable IDs across sequences to avoid double-counting.

Concrete use cases include video summarization where users upload clips for automatic narration; industrial monitoring where systems track objects and events in manufacturing or safety footage; educational applications where students analyze scientific phenomena across multiple images or video sequences; content moderation that identifies specific events or objects in video streams; and research applications where scientists need to ground answers in visual evidence across temporal sequences. The models handle referring expressions like 'Find the window above the kitchen sink,' 'Identify the voice-capturing device held by the woman in yellow,' or 'Locate the animal that makes the plank tip downward' by resolving these queries and returning approximate locations and times.

Molmo 2 is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines and serves a wide audience from practitioners to researchers. The models are available through the Ai2 Playground with video and multi-image workflows, Hugging Face for download, and soon via API. The training code is released under an open-source license, and all newly introduced Molmo 2 open datasets are available with detailed data recipes, plus benchmarks and tools for grounded video evaluation. The architecture scales gracefully at inference time, allowing users to process more frames directly for maximum fidelity or adopt a SlowFast-style strategy that uses high resolution on key frames and lower resolution on others while maintaining similar accuracy on long-video tasks with significantly fewer vision tokens.

Molmo 2 represents a significant advancement in open video understanding, bringing state-of-the-art video pointing, tracking, and dense captioning to the same level of accessibility that Molmo achieved for image understanding. By providing open weights, training data, and training code alongside strong performance across multiple benchmarks, Molmo 2 enables researchers and developers to build systems that anyone can reuse, customize, and improve for multimodal intelligence applications.