Qwen3-VL Release Notes

Ollama has announced the availability of Qwen3-VL on its cloud platform, with local availability planned for a future release. Qwen3-VL is the most powerful vision language model in the Qwen series, designed to integrate high-level visual recognition with advanced reasoning and agentic capabilities.

Advanced Visual and Agentic Capabilities

Qwen3-VL functions as a visual agent capable of operating PC and mobile GUIs by recognizing elements, understanding functions, and invoking tools to complete tasks. It also provides a "Visual Coding Boost," allowing the model to generate HTML, CSS, JS, and Draw.io files directly from images or videos.

Spatial Perception and Grounding

The model features advanced spatial perception, enabling it to judge object positions, viewpoints, and occlusions. This provides stronger 2D grounding and introduces 3D grounding capabilities, which are essential for spatial reasoning and embodied AI applications.

Multimodal Reasoning and Recognition

Qwen3-VL excels in STEM and mathematics, providing causal analysis and evidence-based logical answers. Its visual recognition has been upgraded through broader, higher-quality pre-training, allowing it to identify a wider array of objects, including landmarks, flora, fauna, celebrities, and products.

Document Understanding and Long Context

Qwen3-VL provides seamless text-vision fusion, ensuring that text understanding remains on par with pure Large Language Models (LLMs).

Expanded OCR and Structure Parsing

The model's Optical Character Recognition (OCR) capabilities have expanded to support 32 languages, up from 19. It is designed to be robust against blur, tilt, and low-light conditions, and shows improved performance with jargon and rare or ancient characters. Additionally, the model has improved its ability to parse the structure of long documents.

Video and Long-Context Handling

Qwen3-VL features a native context window of 256K tokens, which is expandable to 1M tokens. This allows the model to process hours-long videos and entire books with second-level indexing and full recall.

Deployment and Integration

Qwen3-VL is currently available via Ollama's cloud service. Users can access the 235B parameter version of the model using the following methods:

Command Line Interface (CLI)

Users can run the model directly using the command:

ollama run qwen3-vl:235b-cloud

Software Libraries

Ollama provides official JavaScript and Python libraries for integration. Both libraries support non-streaming and streaming outputs. For example, using the Python library, users can pull the model and chat with it as follows:

from ollama import chat

response = chat(
    model='qwen3-vl:235b-cloud', 
    messages=[{
        'role': 'user',
        'content': 'What is this?',
        'images': ['./image.jpg']
    }],
)
print(response.message.content)

API Access

The model is accessible via the Ollama cloud API and an OpenAI-compatible API. The OpenAI-compatible endpoints support chat completions, completions, and embeddings, using https://ollama.com/v1 as the base URL.

Sources

Related