Qwen3 Embedding

The Qwen3 Embedding model series offers advanced, versatile solutions for representing and ranking text data, designed to enhance your applications across numerous languages and tasks. Built upon the robust Qwen3 foundational models, this series provides a comprehensive suite of embedding and reranking models optimized for performance and flexibility, helping you unlock deeper text understanding and improve relevance in your systems.

Key Features

• 🌍 Exceptional Multilingual Capability: Leverage support for over 100 languages, including various programming languages. This enables robust multilingual, cross-lingual, and code retrieval capabilities, allowing you to process diverse global text data effectively.

• 📊 State-of-the-Art Performance: Achieve leading results across a wide range of downstream applications. The Qwen3-Embedding-8B model notably ranks #1 on the MTEB multilingual leaderboard (as of June 5, 2025, score 70.58), demonstrating its superior ability to represent text semantics for diverse tasks. The reranking models also excel in improving relevance in text retrieval scenarios.

• ⚙️ Flexible Architecture & Customization: Choose from a spectrum of model sizes (0.6B, 4B, 8B) for both embedding and reranking, balancing efficiency and effectiveness for your specific needs. The embedding models support user-defined output dimensions (from 32 to 4096), allowing you to optimize application costs.

• 🎯 Instruction Awareness: Enhance model performance for specific tasks, languages, or scenarios by providing user-defined instructions. Evaluations show using instructions can typically yield a 1% to 5% improvement, allowing you to fine-tune model behavior for optimal results.

How Qwen3 Embedding Solves Your Problems

Effectively understanding and comparing text is crucial for many applications. The Qwen3 Embedding series provides the tools to transform text into meaningful numerical representations (embeddings) and to accurately reorder search results based on relevance (reranking).

• Building Advanced Search & Retrieval: Create highly relevant search engines or document retrieval systems that work across languages and content types, including code. The embedding models efficiently capture text meaning, while the reranking models refine results for pinpoint accuracy.

• Enhancing Text Classification & Clustering: Improve the performance of your text classification and clustering models by using high-quality, semantically rich embeddings that better capture the nuances of your data, even for multilingual datasets.

• Enabling Bitext Mining & Cross-Lingual Applications: Accurately identify parallel sentences or align text across different languages, facilitating tasks like machine translation training data preparation or cross-lingual information retrieval.

Why Choose Qwen3 Embedding?

The Qwen3 Embedding series offers a unique combination of top-tier multilingual performance, flexible model sizing, and powerful customization features like instruction awareness and dimension control. This empowers you to build high-performing, cost-effective, and globally-aware text understanding applications.

Conclusion

The Qwen3 Embedding model series provides the robust, flexible, and multilingual capabilities you need to tackle complex text representation and ranking challenges. By leveraging its state-of-the-art performance and customizable architecture, you can build more intelligent and effective applications across a wide range of domains.

Explore how Qwen3 Embedding can elevate your text-based applications.

FAQ

• What is the difference between the Embedding and Reranking models?

• Embedding models take a single piece of text (like a sentence or paragraph) and convert it into a fixed-size numerical vector (an embedding) that captures its semantic meaning. These vectors can then be used for tasks like similarity search, clustering, or input into other machine learning models.

• Reranking models take a pair of texts (like a user query and a potential search result) and output a relevance score, indicating how closely related they are. They are typically used after an initial retrieval step to refine the order of results, improving accuracy.

• How does Instruction Awareness work?

• Instruction awareness allows you to prepend a specific instruction to your input text (e.g., "Represent the following sentence for retrieval:" or "Identify the main topic of this text:"). This guides the model to generate embeddings or scores that are better tailored to the specific task or context you require, often leading to improved performance. While the models support many languages, we recommend writing instructions in English as most training instructions were in English.

• What are the different model sizes good for?

• The series offers 0.6B, 4B, and 8B parameter models. Larger models (like 8B) generally offer higher performance and better semantic understanding, suitable for tasks where accuracy is paramount. Smaller models (like 0.6B and 4B) provide a balance of performance and efficiency, making them suitable for applications where computational resources or inference speed are critical constraints. You can choose the size that best fits your specific performance, latency, and cost requirements.