LLaVA-OneVision-2: An open 8B model for video and spatial understanding

EvolvingLMMs Lab released an 8B model that unifies image, long-form video, and spatial understanding. Codec-stream tokenization and windowed attention let it process media at native resolution. Versus Qwen3-VL-8B, it gained 4.3 points on average in video, 5.3 in spatial reasoning, and 15.6 in object-tracking J&F. It trained on about 8M video captions and 4M spatial examples. The full pipeline is available under CC BY 4.0.

Visit original source