LLaVA-OneVision 7B

LLaVA-OneVision 7B is the Qwen 2 7B-based flagship of the LLaVA-OneVision family, released September 2024 by the lmms-lab community. It was designed as a single model for single-image, multi-image, and video scenarios, using a SigLIP vision encoder connected to the language backbone. The 7B variant is the strongest open configuration and remains a widely used baseline for open vision-language research.

Checkpoints are distributed openly (the HF organization llava-hf hosts conversion-friendly builds), and the model has seen broad community adoption including GGUF conversions for llama.cpp and Ascend NPU ports. It is a practical choice for developers who want one open model covering image QA, multi-image comparison, and short video understanding.