Qwen3-Embedding
Alibaba's Apache 2.0 text embedding models in 0.6B, 4B and 8B sizes: 100+ languages, 32K context, adjustable vector length.
In short. Qwen3-Embedding turns text into numbers that capture its meaning, so software can search by meaning rather than keywords. It comes in three sizes: a tiny one that runs almost anywhere and an 8B one for the best accuracy. It handles more than 100 languages, including programming languages, and reads long passages of up to 32,000 tokens.
Summary
Alibaba's Qwen team released the Qwen3 Embedding series on 5 June 2025 under the Apache 2.0 license. The embedding models come in three sizes, all built on the dense Qwen3 language models: 0.6B (28 layers, 1,024-dimensional vectors), 4B (36 layers, 2,560 dimensions) and 8B (36 layers, 4,096 dimensions). Each handles inputs of up to 32,000 tokens and more than 100 languages, including programming languages, so it covers multilingual, cross-lingual and code retrieval.
Two features make them flexible. The vector length can be chosen freely (for the 8B model from 32 up to 4,096 dimensions) to trade accuracy for storage, and the models are instruction-aware: you prepend a short task description to the query, which Qwen says typically improves results by 1% to 5%, and recommends writing instructions in English even for other languages. The series also includes matching Qwen3-Reranker models in the same three sizes, which re-score the best results from the embedding stage.
On release, Qwen reported that the 8B model ranked first on the MTEB multilingual leaderboard with a score of 70.58 as of 5 June 2025; leaderboards move quickly, so treat that as a snapshot and check the current ranking. The BF16 checkpoints are about 15 GB for 8B, 8 GB for 4B and 1.2 GB for 0.6B. For images, screenshots and video, Qwen published a separate family in January 2026, Qwen3-VL-Embedding (2B and 8B) with matching rerankers, which accepts mixed text and visual input in more than 30 languages.
Good for
- Multilingual semantic search and retrieval-augmented generation
- Code search and cross-language retrieval
- Choosing your own trade-off between accuracy and cost, from a 0.6B model for small servers to 8B for best quality
- Two-stage search with the matching Qwen3-Reranker models
Watch out for
- Text only; for images or video use Qwen3-VL-Embedding or EmbeddingGemma 2
- Vectors are not interchangeable between models or sizes, so changing one means re-indexing
- Use the query instruction format and keep it in English for the best results
- The leaderboard ranking is from release time and has likely moved
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Qwen3-Embedding. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Embedding-0.6B BF16 | 600M | 1.2 GB | 1.38 GB | Laptop or phone |
| Embedding-4B BF16 | 4.02B | 8 GB | 9.2 GB | Laptop or phone |
| Embedding-8B BF16 | 7.57B | 15 GB | 17 GB | Gaming GPU |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Qwen3-Embedding vs EmbeddingGemma 2: which should you use?
- Best open-source embedding models in 2026
- EmbeddingGemma 2 embeds text, images, video and audio under Apache 2.0
Frequently asked questions
- What is Qwen3-Embedding?
- Qwen3-Embedding turns text into numbers that capture its meaning, so software can search by meaning rather than keywords. It comes in three sizes: a tiny one that runs almost anywhere and an 8B one for the best accuracy. It handles more than 100 languages, including programming languages, and reads long passages of up to 32,000 tokens.
- Can I use Qwen3-Embedding commercially?
- Yes. Qwen3-Embedding is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Qwen3-Embedding open source?
- Qwen3-Embedding is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Qwen3-Embedding on my own computer?
- Yes. The smallest variant, Embedding-0.6B, needs about 1.38 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Qwen3-Embedding released?
- Qwen3-Embedding was released by Alibaba Qwen on June 5, 2025.
- How long a context window does Qwen3-Embedding have?
- Qwen3-Embedding supports a context window of up to 32K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Qwen blog: Qwen3 Embedding qwenlm.github.io
- Qwen3-Embedding-8B model card (Hugging Face) huggingface.co
- Qwen3-VL-Embedding-8B model card (Hugging Face) huggingface.co
- Qwen3-Embedding repository (GitHub) github.com
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Harrier OSS v1 Microsoft | 270M – 27B | 32K | MIT (Permissive) | 0.621 GB Laptop or phone | |
| EmbeddingGemma 2 Google DeepMind | 270M – 740M | 8K | Apache 2.0 (Permissive) | 0.621 GB Laptop or phone | |
| Qwen3-VL-Embedding Alibaba Qwen | 2.13B – 8.14B | 32K | Apache 2.0 (Permissive) | 4.9 GB Laptop or phone | |
| BGE-M3 BAAI | 568M | 8K | MIT (Permissive) | 2.61 GB Laptop or phone |