New model - Qwen3 Embedding + Reranker
Posted by koc_Z3@reddit | LocalLLaMA | View on Reddit | 3 comments
OP: https://www.reddit.com/r/Qwen_AI/comments/1l4qvhe/new_model_qwen3_embedding_reranker/
Qwen Team has launched a new set of AI models, Qwen3 Embedding and Qwen3 Reranker , it is designed for text embedding, search, and reranking.
How It Works
Embedding models convert text into vectors for search. Reranking models take a question and a document and score how well they match. The models are trained in multiple stages using AI-generated training data to improve performance.
What’s Special
Qwen3 Embedding achieves top performance in search and ranking tasks across many languages. The largest model, 8B, ranks number one on the MTEB multilingual leaderboard. It works well with both natural language and code. Developers aims to support text & images in the future.
Model Sizes Available
Models are available in 0.6B / 4B / 8B versions, supports multilingual and code-related task. Developers can customize instructions and embedding sizes.
Opensource
The models are available on GitHub, Hugging Face, and ModelScope under the Apache 2.0 license.
Qwen Blog for more details: https://qwenlm.github.io/blog/qwen3-embedding/
KittyPigeon@reddit
I suppose using the 8b embedding will increase compute time for a RAG application, but if accuracy matters, it is certainly a much needed and a welcome addition.
Not sure how the reranker model is to be used.
Iory1998@reddit
I think one can use a smaller embedding Qwen-3 model for retrieval and use a larger reranker for accuracy.
--Tintin@reddit
Has anyone used it for RAG already?