Qwen2.5-VL-7B-Instruct

Qwen2.5-VL-7B-Instruct is Alibaba Cloud's Qwen team's open-source instruction-tuned vision-language model, released January 26, 2025 under Apache 2.0. Built on the Qwen2.5-7B LLM backbone with a native dynamic-resolution ViT vision encoder, it accepts text, image, and video inputs and excels at document OCR, chart understanding, visual grounding (bounding boxes and points), structured JSON output, and GUI/computer-use agent tasks. The architecture combines window-attention ViT (SwiGLU + RMSNorm) with Multimodal RoPE (MRoPE) for spatial-temporal alignment, supporting native 32K context extendable to 128K via YaRN.

On official multimodal benchmarks, Qwen2.5-VL-7B leads its size class: MMMU 58.6, DocVQA 95.7, ChartQA 87.3, OCRBench 864, MathVista 68.2, MMStar 63.9, and Video-MME 65.1/71.6 (w/o/w subs). It also scores 84.7 on ScreenSpot and 93.7 on Android Control for agentic UI tasks. On pure text tasks, independent lm-eval testing reports MMLU 71.09 (5-shot); the model preserves core Qwen2.5-7B-Instruct language competencies (HumanEval 84.8, GSM8K 91.6 per the Qwen2.5 technical report). With over 8 million monthly HuggingFace downloads, it is widely regarded as one of the strongest open 7B multimodal models, competitive with GPT-4o-mini on many vision tasks while remaining locally deployable.