Skip to content

Qwen/Qwen3-Embedding-0.6B

Model Information

Qwen/Qwen3-Embedding-0.6B is the smallest model in the Qwen3 Embedding family, optimized for text embedding tasks such as retrieval, clustering, and classification. Despite its lightweight size, it inherits the strong multilingual, long-context, and reasoning capabilities of the Qwen3 foundation models. It supports instruction-aware embeddings, meaning developers can provide task-specific instructions to improve performance (typically a 1–5% boost).

  • Model Developer: Qwen Team (Alibaba Group)
  • Model Release Date: June 2025
  • Supported Languages: 100+ natural and programming languages

Model Architecture

Qwen/Qwen3-Embedding-0.6B model is built on the Qwen3 foundation architecture, which follows a dense Transformer decoder design optimized for embedding tasks. It combines modern architectural components with embedding-specific adaptations:

  • Model Type: Text Embedding
  • Parameters: 0.6B
  • Layers: 28
  • Context Length: 32K
  • Embedding Dimension: Up to 1024 (supports user-defined range 32–1024)
  • Instruction Aware: Yes — queries can include task-specific instructions for better downstream performance
  • MRL Support: Yes — enables flexible embedding vector dimensions for efficiency/accuracy trade-offs
  • Training Corpus: Multilingual + code corpus spanning 100+ languages

Benchmark Scores

Category Benchmark Metric Qwen3-Embedding-0.6B
Multilingual (MTEB) Mean (Task) Avg. Score 64.33
Multilingual (MTEB) Mean (Type) Avg. Score 56.00
Bitext Mining MTEB F1 72.22
Classification MTEB Accuracy 66.83
Clustering MTEB NMI 52.33
Instruction Retr. MTEB Accuracy 5.09
Multilingual Class. MTEB Accuracy 24.59
Pair Classification MTEB Accuracy 80.83
Reranking MTEB MAP 64.64
Retrieval MTEB nDCG@10 61.41
STS MTEB Spearman 76.17

The model provides competitive multilingual embedding performance for its size, outperforming other models in the same parameter class and offering strong flexibility for task-specific customization.

References