Qwen3 Embedding: Advancing Text Embedding and Reranking with Foundation Models

The Qwen3 Embedding series is a new model family specialized for text embedding, retrieval, and reranking, built on top of the Qwen3 foundation model. Leveraging strong multilingual understanding, it aims for SOTA-level performance across multiple benchmarks, and has been released on Hugging Face and ModelScope under the Apache 2.0 license.

Both the embedding and reranker models are provided in 0.6B, 4B, and 8B sizes. The embedding models support vector dimensions of 1024 / 2560 / 4096, along with 32K sequence length, MRL Support, and Instruction Aware capability.

In terms of performance, the reranking models showed strong results on MTEB-family retrieval benchmarks. For example, Qwen3-Reranker-4B recorded MTEB-R 69.76 and FollowIR 14.84, while Qwen3-Reranker-8B achieved the best performance with CMTEB-R 77.45, MMTEB-R 72.94, and MLDR 70.19.

Key features are as follows.

• Exceptional versatility: As of June 5, 2025, the 8B embedding model ranked #1 (70.58) on the MTEB multilingual leaderboard.
• Comprehensive flexibility: Both embedding and reranking are offered in various sizes, and can be adapted to specific tasks, languages, and scenarios via custom instructions.
• Multilingual capability: Supports 100+ languages as well as programming languages, and performs strongly in multilingual, cross-lingual, and code retrieval.

The architecture uses dual-encoder and cross-encoder designs. The embedding model takes a single text as input and uses the hidden state of the final [EOS] token as the representative vector, while the reranker takes text pairs such as a query and a document as input and computes a relevance score.

Training follows the multi-stage approach used in the GTE-Qwen series. The embedding model was trained in 3 stages: contrastive pre-training on weakly supervised data, supervised training on high-quality labeled data, and a merging strategy that combines candidate models. The reranker was trained directly with supervised training on high-quality labeled data.

Notably, in stage 1, a prompt system was introduced that leverages Qwen3's text generation capability to dynamically generate weakly supervised text pairs by task and language. This reduced the limitations of previous approaches that relied on community forums or public data, while efficiently producing large-scale training data.

Going forward, the plan is to further optimize the Qwen foundation model to improve training efficiency and deployment performance for embedding and reranking, and to expand into multimodal representation, broadening the scope to include cross-modal semantic understanding.