What it is
Qwen3-Embedding is a family of open text embedding models from Alibaba's Qwen team, released on 5 June 2025 under Apache 2.0 in 0.6B, 4B and 8B sizes, with matching Qwen3-Reranker models. A reranker rescores a shortlist of search results by reading each query and document together, which is slower but more precise than comparing vectors.
All sizes read up to 32K tokens and support more than 100 natural and programming languages. Output vectors run up to 1,024, 2,560 or 4,096 dimensions by size and can be cut to as few as 32, and both model types accept task instructions, which Qwen says improve results by 1% to 5%. The 0.6B model recorded about 8.3 million Hugging Face downloads in the past month and the 8B about 2 million.
Our review
Qwen3-Embedding remains the open text embedding family to start with, mainly because it covers both stages of a search pipeline under Apache 2.0. Qwen reports 70.58 on MTEB multilingual for the 8B, first on the leaderboard as of 5 June 2025, and 64.33 for the 0.6B; in the same table OpenAI's text-embedding-3-large scored 58.93 and Cohere's multilingual v3 61.12. Those rival figures were taken in May 2025, and both vendors have shipped newer models since.
The reranker is the underrated half. Rescoring the 0.6B embedder's top 100 results, Qwen3-Reranker-4B lifts English retrieval from 61.82 to 69.76 and code retrieval from 75.41 to 81.20, ahead of the BGE, Jina and GTE rerankers in Qwen's comparison. The 4B comes within 1.5 points of the 8B on every row and beats it on two, which makes it the sensible default.
Two usage details decide the results. Queries need a one-sentence task instruction, written in English even for other languages, while documents do not; Qwen says leaving instructions out costs 1% to 5%. Vectors can be cut from 4,096 dimensions to as few as 32, which shrinks a vector database at an accuracy cost Qwen does not quantify.
It is text-only. For images and scanned pages Qwen ships the separate Qwen3-VL-Embedding, which scores 67.88 on MTEB multilingual against this 8B's 70.58, so mixed-media search costs about three points on text. Teams that would rather not run GPUs can use Qwen Cloud's text-embedding-v4, built on Qwen3, at $0.07 per million tokens, a little over half voyage-4-large's $0.12.
Where it earns its place
A search engineer at a developer-tools company indexing 2 million code files
Add semantic code search without per-token fees
Embed with the 0.6B on a single GPU, then rerank the top 100 with Qwen3-Reranker-4B, which Qwen reports lifting MTEB-Code from 75.41 to 81.20. The cost is GPU time rather than an API bill, and the reranking step adds latency to every query.
A product manager at a European marketplace with listings in 12 languages
Improve cross-language search relevance
Run the 8B with English task instructions on queries; Qwen reports 70.58 on MTEB multilingual. The weights are about 16 GB in BF16 by our arithmetic, so one 24 GB GPU serves it and listings never leave your infrastructure.
A start-up CTO with no GPU budget
Embed 5 billion tokens of help articles and tickets
Qwen Cloud's text-embedding-v4 costs about $350 for that volume at $0.07 per million tokens, against about $600 on voyage-4-large. Its limits of 1 million tokens and 1,000 requests a minute mean a backfill that size takes several days.
Before you commit
- Setup
- pip install sentence-transformers (with transformers 4.51 or later) and load Qwen/Qwen3-Embedding-0.6B; vLLM 0.8.5+ and Text Embeddings Inference, including a CPU image, are also supported.
- Integrations
- Sentence Transformers, Transformers, vLLM and Hugging Face Text Embeddings Inference; any vector database that stores float vectors of up to 4,096 dimensions. Hosted text-embedding-v4 on Qwen Cloud.
- Data and privacy
- Apache 2.0 with no usage restrictions; self-hosting keeps documents on your own hardware.
- Cost at scale
- Weights are about 1.2 GB for the 0.6B and 16 GB for the 8B in BF16 by our arithmetic, so the small model runs on almost any GPU and the 8B on one 24 GB card. Hosted text-embedding-v4 costs $0.07 per million tokens.
- Limits
- 32K-token input; vectors up to 1,024, 2,560 or 4,096 dimensions by size. Text only. Rerankers read each query and document together, so rescoring 100 candidates means 100 forward passes per query.
What's good
- 70.58 on MTEB multilingual for the 8B, per Qwen
- Matching rerankers in three sizes; the 4B is close to the 8B
- 32K-token input, 100+ languages, Apache 2.0
- Output dimensions adjustable from 32 to 4,096
What's not
- Benchmark lead dates from June 2025
- Text-only
- Queries need task instructions to reach reported scores
- 8B needs about 16 GB of GPU memory for weights alone
Best for
Teams self-hosting multilingual or code search who want both embedding and reranking under one open licence.
Alternatives
Qwen3-Embedding and Qwen3-Reranker: common questions
- Is Qwen3-Embedding and Qwen3-Reranker worth it?
- A complete open search stack under one licence, embedder and reranker together, though its leaderboard crown dates from June 2025. We rate it 4.2 out of 5.
- How much does Qwen3-Embedding and Qwen3-Reranker cost?
- Free under Apache 2.0. Qwen Cloud's hosted text-embedding-v4, built on Qwen3, is $0.07 per 1M tokens.
- What are the downsides of Qwen3-Embedding and Qwen3-Reranker?
- Benchmark lead dates from June 2025; text-only; queries need task instructions to reach reported scores; 8B needs about 16 GB of GPU memory for weights alone.
- Who is Qwen3-Embedding and Qwen3-Reranker best for?
- Teams self-hosting multilingual or code search who want both embedding and reranking under one open licence.
- How much does Qwen3-Embedding and Qwen3-Reranker cost for a team or in production?
- Weights are about 1.2 GB for the 0.6B and 16 GB for the 8B in BF16 by our arithmetic, so the small model runs on almost any GPU and the 8B on one 24 GB card. Hosted text-embedding-v4 costs $0.07 per million tokens.
- Is Qwen3-Embedding and Qwen3-Reranker safe for company data?
- Apache 2.0 with no usage restrictions; self-hosting keeps documents on your own hardware.
- What are the best Qwen3-Embedding and Qwen3-Reranker alternatives?
- The closest alternatives we have reviewed are voyage-4-large, Cohere Embed 5 and EmbeddingGemma 2.
Answers checked Oct 22, 2026.
Make Qwen3-Embedding and Qwen3-Reranker? Show this rating on your site
