Qwen3 Embedding 0.6B

• 🧠 Smallest model in Alibaba's Qwen3 Embedding series, ~0.6B parameters.
• 📏 Handles context up to 32K tokens, embeddings up to 1024 dimensions.
• 🌐 Multilingual support spanning over 100 languages and code.
• 🔧 Built on dense Qwen3 foundation models for retrieval, clustering, classification.
• 🎯 Instruction-aware with flexible Matryoshka-style dimension representation.
• 🔒 Released under permissive Apache 2.0 license.
• ⚡ Lightweight footprint suited to efficient, low-latency deployments.
• 📚 Companion reranking models available across the same family.

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 0.6B is the entry-level member of Alibaba's Qwen3 Embedding series, a family purpose-built for text embedding and ranking tasks. Built atop the dense Qwen3 foundation models, it inherits their multilingual, long-context, and reasoning capabilities while keeping a compact ~0.6B-parameter footprint. It supports context lengths up to 32K tokens and can emit embeddings with dimensions up to 1024, with flexible Matryoshka-style dimension representation and customizable instruction prompts for tuning retrieval behavior.

The series ships in three sizes, and 0.6B sits below its larger siblings, including Qwen3 Embedding 8B. Per the Qwen3 Embedding technical report, the larger 4B and 8B variants post the strongest MTEB, CMTEB, and code-retrieval scores, while the 0.6B model trades some accuracy for speed and lower memory use. That makes 0.6B the efficiency-oriented option for semantic search, recommendation, and document classification where latency and cost matter most.

For developers, the model integrates with sentence-transformers and is distributed in formats including GGUF for varied hardware. It also pairs with companion Qwen3 Reranker models, letting teams combine dense embedding retrieval with reranking in a single pipeline.

Multilingual coverage spans over 100 languages, with Alibaba advising English-written instructions since most training instructions were originally in English. Apache 2.0 licensing and broad ecosystem support make it accessible for self-hosted retrieval-augmented generation and vector-database workloads.