Qwen3 Embedding 8B

• 🧠 Largest of Alibaba's Qwen3 text-embedding series, 8B parameters
• 📏 Supports 32K-token context, embedding dimensions up to 4096
• 🌐 Multilingual coverage spanning over 100 languages
• 🎯 Built for retrieval, clustering, classification, and reranking
• 🔧 Customizable instructions and user-defined output dimensions
• 🔒 Apache 2.0 license permitting unrestricted commercial use
• 📚 On vendor-reported MTEB multilingual, scores 70.58 (June 2025)
• 🆕 Inherits long-text and reasoning skills from Qwen3 foundation models

Alibaba

Alibaba Group is a Chinese multinational technology company founded in 1999 and headquartered in Hangzhou, Zhejiang. Originally built around e-commerce and cloud computing, Alibaba has become one of the most prolific contributors to open-weight AI research,…

Qwen3 Embedding 8B is the flagship-size member of Alibaba's Qwen3 Embedding series, a family specifically designed for text embedding and ranking tasks. Built on the dense foundational models of the Qwen3 series, it inherits their multilingual capabilities, long-text understanding, and reasoning skills, and supports over 100 languages. The model handles a 32K-token context, produces vectors of up to 4096 dimensions with user-defined output sizes, and is released under the Apache 2.0 license for commercial use.

Within the same family, it sits alongside the much smaller Qwen3 Embedding 0.6B, plus a mid-tier 4B variant. The series spans 0.6B to 8B parameters so developers can trade efficiency against accuracy. The 8B model is the highest-capacity option, intended for workloads where retrieval quality matters most, while the 0.6B is suited to latency- or resource-constrained deployments.

On the provider's reported figures, the 8B embedding model scored 70.58 on MTEB multilingual as of June 5, 2025. The series targets text retrieval, code retrieval, classification, clustering, and bitext mining.

Technical details and training methodology are documented in the team's arXiv paper, "Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models" (2506.05176). The model is widely distributed, with millions of downloads on Hugging Face, and integrates through sentence-transformers and standard embedding-inference tooling.