How Qwen3 Embedding Redefines Multilingual Vector Search Performance

Introduction

In the era of rapid AI development, vectorization has become the foundation of modern AI applications, from search engines to recommendation systems, document retrieval, and semantic analysis. In June 2025, Alibaba's Qwen team released the Qwen3 Embedding series, achieving breakthrough results on multiple benchmarks, especially ranking first on the MTEB multilingual leaderboard with a score of 70.58 for the 8B model.

What Is a Vector Model?

Vector models convert text, images, video and other data into vectors in a mathematical space. By measuring distances or angles between vectors, they quantify similarity, enabling precise search, intelligent recommendation, automatic classification, and anomaly detection.

Current State of Chinese Vector Models

Open‑source Chinese vector models are relatively scarce. The BGE series from BAAI has long been the benchmark, but its medium‑scale models struggle with complex scenarios and long‑text processing.

Market Pain Points:

Scale Limitations: Most open‑source Chinese models have limited parameters, hindering deep semantic understanding.

Context Length: Traditional models typically support only 512‑1024 tokens, making long‑document handling difficult.

Task Generalization: Models fine‑tuned for specific domains lose performance when applied cross‑domain.

Multilingual Ability: Chinese‑focused models perform poorly in mixed‑language scenarios.

The release of Qwen3 Embedding fills this gap, especially the 8B model, which offers commercial‑level performance while remaining open source, marking a new development stage for Chinese vector models.

Qwen3 Embedding Model Overview

Architecture Highlights

Qwen3 Embedding builds on the Qwen3 base model and adopts a dual‑encoder and cross‑encoder architecture. LoRA fine‑tuning preserves and enhances the base model's text understanding capabilities.

Technical Highlights:

Multiple Sizes: 0.6B, 4B, and 8B embedding models.

Long‑Text Support: Up to 32K token input length.

Customizable Dimensions: Users can set output vector dimensions.

Instruction Awareness: Supports task‑specific instruction tuning.

Multilingual Capability: Supports over 100 languages and dialects.

Model Specifications

Text Embedding 0.6B – 28 layers, 32K token limit, 1024‑dim vectors.

Text Embedding 4B – 36 layers, 32K token limit, 2560‑dim vectors.

Text Embedding 8B – 36 layers, 32K token limit, 4096‑dim vectors.

Performance Benchmarks

MTEB Multilingual Benchmark

The 8B model achieved a leading average score of 70.58, outperforming other models across tasks such as classification, clustering, retrieval, and semantic similarity.

Chinese Benchmark (C‑MTEB)

On Chinese text embedding benchmarks, the 8B model scored 73.84 on average, surpassing the 0.6B and 4B variants in all evaluated tasks.

Ollama Local Deployment

Installation and Configuration

<code># Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Run Qwen3 Embedding model
ollama run Q78KG/Qwen3-Embedding-8B:latest</code>

Online API

<code>curl https://ai.gitee.com/v1/embeddings \
  --request POST \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "Qwen3-Embedding-8B",
    "input": "",
    "encoding_format": "float",
    "dimensions": 1,
    "user": null
  }'</code>

This article is based on the latest technical documentation and benchmark results of Qwen3 Embedding, aiming to provide developers with a comprehensive technical reference. Feedback and suggestions are welcome through official channels.