Qwen3-Embedding
Qwen3-Embedding is a family of open text embedding models from Alibaba's Qwen team, published on Hugging Face under the Qwen organization. The family comes in 0.6B, 4B and 8B sizes, alongside matching Qwen3 reranker models. The Qwen3-Embedding-8B model card says the 8B model ranked first on the multilingual MTEB leaderboard as of June 5, 2025, which anchors the family's June 2025 release.
Model Specs
| Spec | Qwen3-Embedding-0.6B | Qwen3-Embedding-8B |
|---|---|---|
| Parameters | 0.6B (28 layers) | 8B |
| Context length | 32K tokens | 32K tokens |
| Embedding dimension | Up to 1024, user-defined from 32 to 1024 | Up to 4096, user-defined from 32 to 4096 |
| License | Apache 2.0 | Apache 2.0 |
A 4B model sits between them, according to the family description on the cards. Sources: the 0.6B card and the 8B card.
Key Features
- Flexible dimensions: the cards describe Matryoshka-style user-defined output sizes, so you can trade accuracy for storage.
- Multilingual and code: support for over 100 languages, including programming languages.
- Instruction aware: the 8B card says task-specific instructions are supported, recommends writing them in English, and reports 1% to 5% retrieval gains when using them.
- Vendor benchmarks: according to the 8B card, it scores 70.58 mean on multilingual MTEB, 75.22 on MTEB English v2 and 73.84 on C-MTEB (Chinese). For the 0.6B model the card lists 64.33 multilingual, 70.70 English v2 and 66.33 C-MTEB. These are Qwen's own results from mid-2025.
- Broad tooling: the card lists Sentence Transformers, Transformers 4.51.0 or later, vLLM 0.8.5 or later and Text Embeddings Inference.
Use Cases
- Semantic search and retrieval for RAG, in many languages.
- Code search, since programming languages are covered.
- Clustering and classification where you want to self-host open weights.
- Pairing with a Qwen3 reranker in a two-stage retrieval pipeline.
A limitation: the embedding sizes differ by model, so switching between 0.6B and 8B means re-embedding your corpus and resizing your vector index. Check newer embedding models before you start a new build.
Getting Started
- Pick a size: 0.6B for cheap, fast embedding, or 8B for the highest reported scores.
- Load it with Sentence Transformers or Transformers, using the card's examples, or serve it with vLLM or Text Embeddings Inference.
- Add a short English task instruction to your queries, and compare retrieval quality with and without it on your own data.
- Choose an output dimension, and store it with your index.
FAQ
What sizes does Qwen3-Embedding come in?
0.6B, 4B and 8B, per the model cards.
How long an input can it take?
32K tokens.
What license does it use?
Apache 2.0, according to the cards.
How many languages does it support?
More than 100, including programming languages.
Alternatives
- BGE-M3: a multilingual embedding model from BAAI.
- EmbeddingGemma: Google's compact open embedding model.
- Jina Embeddings v4: Jina AI's multimodal embedding model.
- OpenAI text-embedding-3-large: a hosted embedding model.
- Qwen3-VL Embedding: the Qwen multimodal embedding sibling.
Conclusion
Qwen3-Embedding is a strong open option when you want multilingual retrieval with flexible dimensions and an Apache 2.0 license. Start from the model card for the current instructions, and see more in the models category.
Sources
- Websitehttps://huggingface.co/Qwen
- Social Mediahttps://x.com/Alibaba_Qwen
Comments
No comments yet. Be the first to comment!