Qwen-VL

Alibaba Cloud's first open vision-language model arrived in August 2023 as Qwen-VL and its dialogue-tuned sibling Qwen-VL-Chat, pairing the Qwen-7B language model with a large ViT image encoder fed 448 by 448 pixel inputs. That resolution, plus training aimed at reading text and returning bounding boxes, made it unusually good with documents, charts and grounding questions for its time, and an Int4 build lowered the memory bar. The last commit to the repository was in April 2024, and the model line has since moved on through Qwen2-VL, Qwen2.5-VL and Qwen3-VL. The original weights still load from Hugging Face, but only with trust_remote_code and an old Transformers release, since the repo pins version 4.32.0, which makes it awkward inside current serving stacks. Mind the license: weights and code use the Tongyi Qianwen License Agreement rather than Apache-2.0; it permits commercial use below 100 million monthly active users and forbids using outputs to improve other LLMs. For new work, choose Qwen3-VL, which is Apache-2.0, or Qwen2.5-VL, checking the license of the specific size you pick.