VideoLLaMA 3 drops

Video models try to answer a hard question: not just what is in one frame, but what changes from second to second. VideoLLaMA 3 is a new open model built for that job, with a 7 billion-parameter version posted on Hugging Face on January 21, 2025. (huggingface.co) The release includes a 7B video model, a 2B video model, and image-focused 7B and 2B variants. The 7B checkpoint is based on Qwen2.5-7B, uses an Apache 2.0 license, and is distributed as a video-to-text model on Hugging Face. (huggingface.co) In practice, the model takes sampled frames from a clip and text prompts such as “Describe this video in detail.” The published quick-start example sets video sampling at 1 frame per second and caps the input at 128 frames. (huggingface.co) The underlying problem is temporal reasoning, which means tracking order, motion, and cause across frames instead of treating a video like a pile of still images. The VideoLLaMA 3 paper says the model is tuned to handle both image and video understanding, with a final video-centric fine-tuning stage aimed at improving video capability. (arxiv.org) The paper describes a “vision-centric” approach that leans heavily on large, high-quality image-text data before adding more video-specific training. It also says the model adapts to variable image resolutions and prunes visually similar video tokens so the video representation stays more compact. (arxiv.org) That design reflects a broader shift in multimodal artificial intelligence, where developers are trying to make one model read documents, inspect images, and follow action across clips with the same backbone. The GitHub repository describes VideoLLaMA 3 as a series for image and video understanding rather than a tool built for one narrow benchmark. (github.com) The team tied the launch to benchmark claims within days of release. The GitHub repository says VideoLLaMA3-7B was the best 7B-sized model on the VideoMME leaderboard as of January 24, 2025, and on LVBench as of January 26, 2025. (github.com) VideoLLaMA has been iterating quickly: the original Video-LLaMA repository was published in 2023, and the VideoLLaMA 3 technical report was posted to arXiv on January 22, 2025. By April 2026, the Hugging Face model page was still live as the main public checkpoint for developers who want to run or fine-tune it. (github.com) (arxiv.org) (huggingface.co) For developers, the immediate change is simple: an open 7B model for asking questions about short videos is now packaged with code, weights, and a paper that explains how it was trained. That puts the focus back on whether smaller open models can keep improving on video tasks without the cost of much larger systems. (huggingface.co) (github.com)