QwenLM

Qwen3-VL: Understand Images, Video, and Text with Multimodal Models

Qwen3-VL is a family of multimodal language models for interpreting images, video, and text. The repository provides inference examples, deployment guidance, and cookbooks for tasks such as OCR, spatial reasoning, and visual agents.

Jupyter NotebookApache-2.0AILLMMachine Learning
View on GitHub Compare
Stars
20k
Forks
1.9k
Stars, last 30 days
+137

Our take

Health score

38/100

How it is scored
Activity
19
Community
68
Issues
46
Pull requests
20

+137 stars in the last 30 days

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

20,220

Forks

1,899

Change since Nov 16, 2026: +6 stars·+1 forks

Overview

Qwen3-VL is a vision-language model series from the Qwen team. It combines image, video, and text input to support tasks ranging from visual question answering and document reading to video analysis and spatial reasoning.

The repository is primarily a model-use and deployment resource: it links to model checkpoints and offers examples for running inference with Transformers, vLLM, and SGLang. It is aimed at developers and researchers who need multimodal capabilities, rather than users looking for a small standalone application.

Key Features

  • Processes text alongside images and video, including multiple visual inputs.
  • Provides Instruct and Thinking variants across Dense and MoE model configurations.
  • Supports OCR, document parsing, object grounding, video understanding, and spatial reasoning.
  • Includes cookbook notebooks for visual tasks, multimodal coding, and computer or mobile agent workflows.
  • Documents local inference and serving with Transformers, vLLM, and SGLang.
  • Offers controls for image and video resolution, frame sampling, and visual-token budgets.
  • Describes a 256K-token context configuration and YaRN settings for extending context up to 1M tokens.

Use Cases

  • Researchers can evaluate vision-language models on visual reasoning or multimodal tasks and consult the published evaluation settings.
  • Developers can build document-processing workflows that extract text and layout information from images or long documents.
  • Teams can prototype video analysis applications that summarize clips, read video text, or locate events over time.
  • Agent developers can explore GUI automation workflows that interpret screen content and use tools to interact with computers or mobile devices.

Getting Started

Install the documented Transformers minimum:

pip install "transformers>=4.57.0"

Choose a Qwen3-VL checkpoint from the Hugging Face collection or ModelScope collection. See the README for inference examples, dependencies, and deployment instructions.

Alternatives

  • Ollama-OCR: Ollama-OCR focuses on extracting text and structured content from images and PDFs using Ollama-hosted models, rather than general multimodal reasoning.
ProjectLanguageLicenseStarsStatus
Qwen3-VLJupyter NotebookApache-2.020kNot checked yet
Ollama-OCRJupyter NotebookMIT2.8kInactive

Considerations

  • Model sizes range from edge-oriented configurations to very large checkpoints. Local inference, especially with the 235B MoE model, can require substantial GPU memory and multi-GPU deployment; the README's serving example targets H100 or H200 hardware.
  • Framework compatibility and requirements vary. The README specifies Transformers 4.57.0 or later and vLLM 0.11.0 or later for the documented support.
  • Video processing depends on third-party decoding backends and their compatibility. The repository notes that backend choice affects speed and URL support.
  • Some evaluation benchmarks and reproduction assets are described as internal or planned for later release, so the documented results may not be fully reproducible from this repository alone.

Found this useful?

Share it with someone who would like Qwen3-VL.

Comparisons