Alibaba released core-emb-2b on Hugging Face
Alibaba released core-emb-2b, a 2B multimodal embedding model, on Hugging Face under CC-BY-4.0. It is built on Qwen3-VL and trained via reranker distillation for compositional reasoning. The model is available for download and use with the transformers library.
China context
- Original name
- 阿里巴巴
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can download the model weights from Hugging Face and integrate it into multimodal retrieval systems, with usage requiring a recent transformers build supporting Qwen3-VL.
- For investors
- Alibaba continues to release open-weights models, which may pressure competitors and expand its developer ecosystem.
Core-emb-2b uses Rank-KL distillation to reproduce a reranker's fine-grained ranking over a five-level compositional matching spectrum, improving attribute-object binding. It preserves general retrieval performance on COCO and Flickr30k while improving compositional accuracy.
Developers outside China can now use a 2B open-weights multimodal embedding model that improves compositional retrieval without sacrificing general retrieval quality, reducing the need for custom reranking pipelines.
Businesses can integrate core-emb-2b into multimodal search and retrieval systems to better handle fine-grained queries involving attribute-object bindings, improving accuracy in e-commerce or content moderation.
The release of core-emb-8b and core-reranker models may follow, as the model card lists them in the same family. Adoption can be tracked via Hugging Face downloads and community benchmarks.