Why Qwen3 Embedding Models Are Setting New Benchmarks in Text Representation

Overview of Qwen3 Embedding Series

The Qwen3 Embedding series, released under the Apache 2.0 license on Hugging Face and ModelScope, provides text representation and ranking models designed for retrieval and sorting tasks. It builds on the Qwen3 base model, inheriting strong multilingual understanding and achieving top scores on multiple benchmark leaderboards.

Key Features

Outstanding Generalization: The 8B‑parameter embedding model ranks first on the MTEB multilingual leaderboard (score 70.58 as of June 5 2025), surpassing many commercial APIs.

Flexible Architecture: Three model sizes (0.6B, 4B, 8B) allow trade‑offs between performance and efficiency. Users can customize embedding dimensions and instruction templates for specific tasks.

Comprehensive Multilingual Support: Supports over 100 languages, including programming languages, enabling cross‑language and code retrieval.

Model Overview

Model Type          Models                     Size   Layers  SeqLen  EmbDim  MRL Support  Instruction Aware
Text Embedding      Qwen3-Embedding-0.6B       0.6B   28      32K     1024    Yes          Yes
                    Qwen3-Embedding-4B         4B     36      32K     2560    Yes          Yes
                    Qwen3-Embedding-8B         8B     36      32K     4096    Yes          Yes
Text Reranking     Qwen3-Reranker-0.6B        0.6B   28      32K     -       -            Yes
                    Qwen3-Reranker-4B          4B     36      32K     -       -            Yes
                    Qwen3-Reranker-8B          8B     36      32K     -       -            Yes

Model Architecture

Both embedding and reranking models are built on the Qwen3 foundation. Embedding models use a dual‑tower design: a single text segment is fed to the model, and the hidden state corresponding to the [EOS] token is extracted as the semantic vector. Reranking models employ a single‑tower design that takes a pair of texts (e.g., query and candidate document) and outputs a relevance score.

Training Procedure

The embedding models follow a three‑stage training pipeline inherited from the GTE‑Qwen series: (1) large‑scale weak‑supervised contrastive pre‑training, (2) supervised fine‑tuning on high‑quality annotated data, and (3) model‑fusion of multiple candidates to boost overall performance. This balances generalization and task‑specific adaptation.

Reranking models are directly supervised on high‑quality labeled data, streamlining training. For the first stage of embedding training, a multi‑task prompt system generates diverse weak‑supervised text pairs using Qwen3’s generation capability, eliminating reliance on community‑scraped data.

Evaluation Results

Ranking performance is evaluated using top‑100 vector retrieval from Qwen3‑Embedding‑0.6B followed by reranking. The following metrics are reported (higher is better):

Model                     Param   MTEB‑R   CMTEB‑R   MMTEB‑R   MLDR   MTEB‑Code   FollowIR
Qwen3‑Embedding‑0.6B      0.6B    61.82    71.02    64.64    50.26   75.41      5.09
Jina‑multilingual‑v2‑base 0.3B    58.22    63.37    63.73    39.66   58.98     -0.68
... (additional rows omitted for brevity) ...
Qwen3‑Reranker‑8B        8B      69.02    77.45    72.94    70.19   81.22      8.05

Note: Qwen uses MTEB (eng, v2), MTEB (cmn, v1), MTEB (Multilingual) and MTEB (Code) datasets, denoted as MTEB‑R, CMTEB‑R, MMTEB‑R, and MTEB‑Code respectively. Ranking results are based on top‑100 vectors retrieved by Qwen3‑Embedding‑0.6B.

Future Directions

Qwen3 Embedding will continue to improve training efficiency and deployment performance, and plans to extend into multimodal representation, building cross‑modal semantic understanding. The team encourages developers to explore broader application scenarios using the series.