← All models

Qwen2.5-VL-Embedding

Published Publisher Alibabaqwen2.5-vl-embedding

Positioning

Qwen2.5-VL-Embedding is a unified multimodal embedding model based on Qwen2.5-VL. It converts text, images, video, or mixed-modality inputs into a shared representation vector. Alibaba lists it as released on October 21, 2025. (help.aliyun.com)

Capabilities and uses

The model is intended for cross-modal retrieval, image and video search, clustering, and tagging. It supports embedding dimensions of 512, 768, 1,024, and 2,048; the default is 1,024. (help.aliyun.com)

Limitations

  • It produces fused embeddings, not separate independent vectors for each modality. The documented text limit is 32,000 tokens, and each image must be no larger than 5 MB. (help.aliyun.com)
  • Per request, each of the text, image, video, and fused-object types may appear at most once. The documentation lists support for 11 text languages, including Chinese and English. (help.aliyun.com)
  • The model page lists function calling, structured outputs, web search, context caching, batch inference, and fine-tuning as unsupported. (help.aliyun.com)

Pricing

Alibaba Cloud Model Studio, China (Beijing), standard API pricing: text input is CNY 0.7 per 1 million input tokens; image/video input is CNY 1.8 per 1 million input tokens. Output is free, and the model has no free input quota. These are standard prices, excluding promotions. Source: official mainland pricing, checked November 17, 2026. (help.aliyun.com)

The official international USD pricing page, checked November 17, 2026, does not list a price for this model; an international USD price is therefore unverified. (alibabacloud.com)

Official sources