Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, and Text Recognition

Generalist VLMs of comparable scale lacked grounded text reading and localisation ability.

Qwen-LM foundation plus a designed visual receptor, an input-output interface, a 3-stage training pipeline and a multilingual multimodal cleaned corpus. Grounding is trained by aligning image-caption-BOX tuples rather than free-form captions alone.

Technical report, no peer-reviewed venue found. The Qwen-VL family is one of the most widely deployed open VLMs, and the Qwen2.5-VL line that followed made dynamic-resolution native - images are resized to fit a token budget, which is a serving-visible behaviour: cost tracks the configured resolution, not the source image. The Qwen2-VL successor introduced window attention for large images to keep the vision encoder's own attention cost sub-quadratic; the original Qwen-VL here does not have it. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'unknown'. `venue` is 'arxiv-preprint', which means no peer-reviewed venue has been established for this work as of the verification date, so the track genuinely is not known. 'unknown' is used here rather than leaving the field null or guessing 'main': the schema separates 'unknown' from 'main' precisely so that an unprereviewed preprint is not silently counted as main-track work. This annotation is what lets a downstream filter distinguish 'peer reviewed, main track' from 'not peer reviewed' without inferring it from the venue string.