Qwen3-VL-Embedding

Qwen3-VL-Embedding is a series of models designed for high-precision multimodal retrieval and ranking. Built upon the Qwen3-VL foundation model, it processes diverse inputs—including text, images, screenshots, and videos—by mapping them into a unified representation space. This enables efficient similarity computation and retrieval across different modalities.

Qwen3-VL-Embedding is a series of models designed for high-precision multimodal retrieval and ranking. Built upon the Qwen3-VL foundation model, it processes diverse inputs—including text, images, screenshots, and videos—by mapping them into a unified representation space. This enables efficient similarity computation and retrieval across different modalities.

• 01Supports over 30 languages
• 02Processes text, images, screenshots, and videos
• 03Offers flexible embedding dimensions (64–2048)
• 04Utilizes Matryoshka Representation Learning for customizable embedding sizes
• 05Handles inputs up to 32k tokens