Qwen3-Embedding-4B

Qwen3-Embedding-4B

Turns text and code in over 100 languages into vectors for search, retrieval, classification and clustering, with adjustable vector size and task instructions, for developers building search and RAG systems.

Made by Qwen on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen3-Embedding-4B turns text into vectors, so you can compare meaning by comparing numbers. It has 4B parameters, reads up to 32k tokens and supports over 100 languages, including programming languages. It belongs to the Qwen3 Embedding series, which also has reranking models. The vector size can be set anywhere from 32 to 2560.

Who it suits: Developers who build search, retrieval or recommendation features will get the most from it, as will teams who sort or group large piles of text. It suits multilingual work well.

Industry | Who uses it | How they use it | Example
Software | Search engineers | Embed documents and queries, then rank by closeness | Semantic search over a knowledge base
Support | Support teams | Match a new question to past answers | Suggesting related help articles
Data teams | Analysts | Group similar texts into clusters | Sorting survey comments by theme
Code tools | Developer-tool builders | Retrieve code snippets from a plain-language query | Finding a function by describing it

If you need a model that writes text, you can skip it, since this one only produces vectors.

Where to use it:

• Python with the sentence-transformers library, which Hugging Face lists
• Search and retrieval pipelines that store vectors
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Text retrieval | Search engineers | A vector for each query and document
Text classification | ML engineers | Vectors to feed a classifier
Clustering | Analysts | Vectors that group similar texts
Bitext mining | Translation teams | Vectors to pair matching sentences across languages

How to set it up: You need Python with pip and transformers 4.51.0 or newer. Older versions stop with a KeyError for qwen3. The weights are about 8 GB in BF16.

• Python with pip, on Windows, macOS or Linux
• sentence-transformers
• transformers 4.51.0 or newer
• About 8 GB of space for BF16 weights

1. Install the Hugging Face hub package.
2. Download the model files.
3. Load the model with sentence-transformers.

pip install -U huggingface_hub
hf download Qwen/Qwen3-Embedding-4B

How to use it: Pass your texts to the model through sentence-transformers and it returns one vector for each. You can choose a smaller vector size, and you can add an instruction that describes your task. The source shows no code, so follow the model page for the exact call.

Tool | What it does
sentence-transformers | The library that loads the model and makes the vectors
Custom dimensions | Sets the output vector size from 32 up to 2560
Instructions | Adds task text to the input, which the card says helps by 1 to 5 percent

Example: You embed the question "How do I reset my password" together with a set of help articles. You rank the articles by how close their vectors are to the question and show the top few.

For newcomers and experts: An embedding is a list of numbers that stands for the meaning of a text, so similar texts get similar lists. A good first step is to embed two short sentences and compare them.

The model is built on Qwen3-4B-Base with 36 layers and a 32K sequence length. Its dimension can be shrunk, which the card calls MRL support. It is instruction aware, and the card advises writing instructions in English even for other languages. It pairs with the series rerankers, which can reorder the top results.

You pass in text, plus an optional task instruction
The model reads it with its 36 layers
It returns one vector, up to 2560 dimensions
You compare vectors to rank or group texts

Title: Embedding dimension by model size
0.6B model: 1024 dimensions
4B model: 2560 dimensions
8B model: 4096 dimensions

Good to know:

• License: apache-2.0, as listed on Hugging Face
• Cost to run: none named
• Data it reaches: the text you pass in, and the source names no web or account access
• Supports over 100 languages, including programming languages
• Not for: text generation, because it only produces vectors
• Needs transformers 4.51.0 or newer

On Hugging Face, 13 November 2026

• Task: feature extraction
• Library: sentence-transformers
• Parameters: 4B
• Weights: about 8 GB in BF16; about 2 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,426,700
• Likes: 351
• License: apache-2.0
• Based on: Qwen/Qwen3-4B-Base
• Quantized versions: 53
• Fine-tunes: 77
• Tags: sentence-transformers, safetensors, qwen3, text-generation, transformers, sentence-similarity, feature-extraction, text-embeddings-inference

Links

• Model: https://huggingface.co/Qwen/Qwen3-Embedding-4B
• Files: https://huggingface.co/Qwen/Qwen3-Embedding-4B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Qwen/Qwen3-Embedding-4B

Then follow the usage in its model card. Nothing in this kit runs by itself.