Qwen3-Embedding-8B

Description

The Qwen3 Embedding model by Alibaba is designed for creating embeddings, which are essential for a retrieval-augmented generation (RAG) system. This model does not support chat interaction. It is used to convert documents and texts into individual vectors, which are stored in a vector database by the application, enabling semantic search. The strength of this model lies in its support of over 100 languages with high performance.

The following limits and defaults apply:

• Maximum context length: 32,768 tokens
• Default embedding dimension: 4,096
• Supported values for dimensions: 256, 512, 768, 1,024, 1,536, 2,048, 3,072, and 4,096

It is recommended to format embeddings for search queries using the following template:

Instruct: {task_description}
Query: {query}

Here, {query} should express the individual search query in one sentence, and {task_description} should describe the task, for example:

Reducing embedding dimensions

Qwen3-Embedding-8B supports server-side dimension reduction through the OpenAI-compatible dimensions parameter. This is a supported model capability based on Matryoshka Representation Learning.

Set dimensions to 256, 512, 768, 1,024, 1,536, 2,048, 3,072, or 4,096. The endpoint returns a vector with exactly the requested number of values and L2-normalizes it after reducing its dimensions. Values below 256 are not supported because retrieval quality degrades at smaller dimensions.

For example, request an embedding with 256 dimensions:

from openai import OpenAI

client = OpenAI(
    base_url="https://llm.aihosting.mittwald.de/v1"
)

response = client.embeddings.create(
    model="Qwen3-Embedding-8B",
    input="Testing 123",
    encoding_format="float",
    dimensions=256,
)

print(len(response.data[0].embedding))  # 256

Smaller embeddings reduce transfer and storage requirements but can also affect retrieval quality. Evaluate the chosen dimension with data representative of your use case.

Document RAG pipeline

For a complete end-to-end example — combining Qwen3-Embedding-8B with GLM-OCR for document ingestion and Qwen3.5-122B-A10B-FP8 for answering — see the GLM-OCR guide. It also covers vector database options (pgvector, Qdrant, ChromaDB) on mittwald Container Hosting.

Terms of use and licensing

The general terms of use apply. The model is provided by Alibaba under the Apache 2.0 License, and reuse of the generated content is not subject to any additional restrictions.