Molmo2-8B is an open multimodal vision-language model developed by the Allen Institute for AI (Ai2). Released as part of the Molmo2 family, it is designed for advanced spatial and temporal understanding of images, videos, and multi-image sets. The model aims to bridge the performance gap between open-weight systems and proprietary models in grounded vision tasks.
Architecture and Development
The model is built upon the Qwen3-8B language base and utilizes SigLIP 2 as its vision backbone. It was trained on the Molmo2 data collection, which includes over 9 million highly curated multimodal examples featuring dense video captions, long-form QA, and tracking data. This emphasis on data quality allows the 8B-parameter model to achieve performance parity with significantly larger models, including the original 72B Molmo and various proprietary systems, particularly in video-related benchmarks.
Key Capabilities
Molmo2-8B specializes in video grounding, which involves identifying and tagging objects across temporal frames. It supports variable-length video inputs and excels at tasks such as multi-object tracking, pixel-level grounding, and complex counting. Its multi-image reasoning capabilities enable it to compare and analyze multiple visual inputs within a single conversational context, making it suitable for applications in robotics and automated document analysis.