Qwen Text Embedding v4

Overview

Qwen Text Embedding v4 is a 4B-parameter dual-encoder model from the Qwen3 family, optimized specifically for dense embeddings and ranking tasks rather than general chat. It supports over 100 languages (including major programming languages) and is tuned for semantic search, retrieval, classification, clustering, and bitext mining in a single shared embedding space.

Technical Specifications

• Architecture: Dense transformer-based encoder with symmetric contrastive fine-tuning
• Context Length: 8,192 tokens
• Capabilities: Multilingual semantic encoding, query/document alignment, cross-lingual retrieval, similarity ranking
• Training Data: Curated corpus spanning technical documentation, academic papers, conversational logs, and web-scale multilingual text with rigorous deduplication and bias mitigation

Performance Benchmarks

• MTEB (Massive Text Embedding Benchmark): Achieves top-tier performance among open and closed models, excelling in retrieval, classification, and clustering subtasks
• Multilingual Alignment: Maintains >92% cross-lingual similarity fidelity on aligned sentence pairs across major language families

Quality Improvements

• Semantic Density: Embeddings exhibit tighter intra-cluster cohesion and sharper inter-cluster separation compared to v3
• Noise Resilience: Robust to input perturbations, formatting inconsistencies, and moderate grammatical errors
• Bias Control: Integrated fairness-aware training reduces spurious correlations in gender, region, and domain-sensitive dimensions

API Pricing

• $0.0735 / 1M tokens

New Features & Technical Upgrades

Qwen Text Embedding v4 introduces a suite of innovations focused on semantic fidelity, efficiency, and multilingual equity:

Key Features

• High Semantic Fidelity: Captures fine-grained semantic relationships, even in complex or domain-specific phrasing.
• Long-Context Awareness: Handles inputs up to 8K tokens—ideal for embedding full documents or detailed user queries.
• Multilingual Robustness: Unified embedding space across languages enables cross-lingual retrieval without translation.
• Optimized for Retrieval: Trained with contrastive and in-batch negative sampling for superior performance in similarity search.
• Low Latency & High Throughput: Efficient inference pipeline suitable for real-time applications at scale.

Practical Impact

‍These enhancements translate into stronger RAG accuracy, more coherent document clustering, and reduced false positives in semantic search, especially in multilingual support portals, research knowledge bases, and cross-border enterprise analytics.

Code Sample

Comparison with Other Models

vs OpenAI text-embedding-3-large: Qwen v4 matches or exceeds OpenAI’s performance on MTEB while offering 8K context (vs 8K for OpenAI, but with lower cost and data residency flexibility). Unlike OpenAI, Qwen provides transparent licensing for commercial deployment and avoids data usage for model training.

vs Google’s textembedding-gecko: Qwen v4 provides better zero-shot retrieval scores on BEIR and avoids vendor lock-in through open-weight availability. Gecko integrates tightly with Vertex AI, while Qwen offers greater deployment flexibility across clouds and on-prem.