Qwen2.5-VL is a series of vision-language models launched by the Alibaba Cloud Tongyi Qianwen (Qwen) team. This GitHub repository currently primarily hosts the official documentation and resources for Qwen3-VL, while also retaining the related release information and code for Qwen2.5-VL and Qwen2-VL.
Purpose The series is aimed at multimodal understanding and generation tasks, and can simultaneously process text, image, and video inputs, for use in scenarios such as visual question answering, document parsing, video understanding, OCR, spatial grounding, and reasoning.
Main Capabilities According to the official documentation, Qwen3-VL has been upgraded in the following aspects:
- Visual Agent: can operate PC/mobile GUIs, recognize interface elements, understand functions, call tools, and complete tasks.
- Enhanced visual coding: can generate Draw.io, HTML/CSS/JS from images/videos.
- Spatial perception: judges object position, viewpoint, and occlusion, provides stronger 2D grounding, and supports 3D grounding for spatial reasoning and embodied AI.
- Long context and video understanding: native 256K context, expandable to 1M; supports processing books and multi-hour long videos, with complete recall and second-level indexing.
- Multimodal reasoning: performs outstandingly in STEM/mathematics, supports causal analysis and evidence-based logical answers.
- Visual recognition: through broader and higher-quality pretraining, can recognize celebrities, anime, products, landmarks, animals and plants, etc.
- OCR expansion: supports 32 languages (previously 10), is more robust under low-light, blurry, and tilted conditions, has better recognition of rare/ancient scripts and terminology, and improved parsing of long document structures.
- Text understanding: on par with pure-text LLMs, achieving lossless unified understanding of text and vision. Technical features include Interleaved-MRoPE, DeepStack, and Text-Timestamp Alignment.
Available Versions and Architectures The series provides Dense and MoE architectures, covering scales from edge to cloud, and includes Instruct and Thinking versions with enhanced reasoning. Released versions include Qwen3-VL-235B-A22B, 30B-A3B, 32B, 8B, 4B, 2B, etc. (including Instruct/Thinking and FP8 versions). Qwen2.5-VL provides 3B, 7B, and 72B parameter scales, and has released AWQ quantized models.
Usage The repository provides code examples for calling the models via 🤗 Transformers and ModelScope, and provides the qwen-vl-utils toolkit for controlling visual input processing. Users can control the pixel budget for images and videos by setting parameters such as min_pixels, max_pixels, resized_height, resized_width, total_pixels, etc.; supports local files, base64, and URL inputs. Video decoding supports torchvision, decord, and torchcodec backends, with torchcodec officially recommended. The model can use Flash Attention 2 when loaded (requires float16 or bfloat16). When the context exceeds 256K, it can be extended via YaRN (note that Interleaved-MRoPE requires a smaller scaling factor). For deployment, vLLM is recommended (requires vllm>=0.11.0), SGLang is also supported, and it can be accessed through an OpenAI-style API; an official Docker image qwenllm/qwenvl is provided. The repository also provides a way to launch a web UI demo.
Applicable Scenarios Suitable for research and engineering scenarios requiring multimodal understanding, such as document and video parsing, GUI automation, visual reasoning, multilingual OCR recognition, spatial grounding, embodied AI, etc.
Limitations The materials do not provide specific information on fees, pricing, or performance metrics. Some evaluation benchmarks are internally constructed, and their code and reproduction resources are planned for later release; the evaluation prompts for a few benchmarks have been modified, and details will be explained in the technical report. decord has issues such as decoding hangs, and the project is no longer actively maintained.
