Qwen2.5-VL-7B-Instruct

Qwen2.5-VL-7B-Instruct is the instruction-tuned release of the 7-billion-parameter model from the Qwen2.5-VL family developed by the Qwen Team at Alibaba Group. It is a VLM (Vision-Language Model) designed to process multimodal inputs: text, static images and video. Model weights are publicly available under the Apache 2.0 licence.

Architecture

The model is based on the Transformer architecture with an attached ViT (Vision Transformer) visual encoder. The visual encoder has been optimised using window attention, SwiGLU and RMSNorm — accelerating training and inference while maintaining performance. For video processing, Dynamic FPS Sampling and mRoPE with absolute time alignment are used — enabling understanding of videos of varying lengths and frame rates. The model natively supports a 32,768-token context window (extendable via YaRN). Dynamic Input Resolution: images are processed without altering aspect ratio, and the number of visual tokens adjusts to content (default range 4–16,384 tokens per image).

Key capabilities

The model excels in five areas: (1) document understanding — analysis of charts, tables, invoices, forms and scans with structured JSON output; (2) multi-level OCR — text recognition in scene photos, documents and handwriting; (3) long video understanding — videos over 1 hour with precise event timestamp pinpointing; (4) visual localisation — object detection via bounding box or point with JSON output; (5) agentic capabilities — computer and phone use as a visual agent (ScreenSpot 84.7, AndroidControl Low_EM 93.7).

Availability

The model is available for download from Hugging Face and ModelScope under the Apache 2.0 licence. It supports the Transformers, vLLM and SGLang libraries. Flash Attention 2 is recommended for efficient inference. The model is in BF16 format and requires a GPU with at least ~16 GB VRAM. Dynamic min_pixels/max_pixels configurations are available to balance inference quality and speed.