Qwen3 Embedding

Qwen3 Embedding is a model series from the Qwen team (QwenLM) built on the dense Qwen3 foundation models. It provides text embedding and text reranking models in 0.6B, 4B and 8B sizes, with usage examples for Transformers, vLLM and Sentence Transformers.

What It Is

Qwen3 Embedding converts text into vectors and reranks candidate passages for a query. The README lists target tasks including text retrieval, code retrieval, text classification, text clustering and bitext mining. The models are published on Hugging Face and ModelScope, and an API is linked from the README.

Model Lineup and Flexibility

The embedding models (Qwen3-Embedding-0.6B, 4B, 8B) have 28 to 36 layers, 32K sequence length, and embedding dimensions of 1024, 2560 and 4096. They support MRL, which allows custom output dimensions, and are instruction aware. The reranker models (Qwen3-Reranker-0.6B, 4B, 8B) also accept user-defined instructions. The README says embedding and reranking modules can be combined, for example by reranking the top candidates retrieved by a dense embedding model.

Multilingual and Code Support

The series inherits Qwen3's support for over 100 languages, including various programming languages, covering multilingual, cross-lingual and code retrieval.

Usage Notes

Transformers 4.51.0 or later is required, and vLLM 0.8.5 or later for the vLLM examples. Queries should be prefixed with a one-sentence task instruction, while documents need none. The README says instructions typically improve results by 1% to 5% and recommends writing them in English. Flash attention 2 is suggested for acceleration and memory savings.

Evaluation Claims

The Qwen team reports that the 8B embedding model ranked No.1 on the MTEB multilingual leaderboard as of June 5, 2025, with a score of 70.58. The README also provides tables for MTEB English v2, C-MTEB and reranking benchmarks, and points to training docs in the repository.

License and costs

The official 0.6B and 8B embedding model cards identify Apache-2.0 licensing, despite the GitHub README’s inconsistent “proprietary” wording. Downloadable models and local execution should be distinguished from the separately linked Alibaba Cloud API; no hosted API price is asserted here. Check the exact selected model artifact’s terms and budget local compute separately.