bge-small-zh-v1.5

bge-small-zh-v1.5

Turns Chinese text into dense vectors for search, classification, clustering and semantic matching. A 24M-parameter BGE embedding model from BAAI under the MIT license, about 96 MB in F32.

Made by BAAI 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: bge-small-zh-v1.5 is a small Chinese text embedding model from BAAI, part of the FlagEmbedding project. It maps a piece of text to a short list of numbers, called a dense vector. These vectors serve retrieval, classification, clustering and semantic search, and they can sit in the vector databases that LLM apps use. Version 1.5 was released to give a more reasonable spread of similarity scores and better retrieval without an instruction prefix.

Who it suits: It suits developers who build search or retrieval over Chinese text and want a model that is small to download and quick to run.

Industry | Who uses it | How they use it | Example
Search | Teams building document search | Embed documents and queries, then compare the vectors | Find Chinese passages related to a question
Customer support | Teams matching questions to answers | Embed saved answers and look up the closest one | Match a customer's question to a stored reply
AI apps | Developers building retrieval for LLMs | Store the vectors in a vector database | Fetch relevant Chinese notes before asking a model

You can skip it if your text is not Chinese, since Hugging Face lists only Chinese for this model.

Where to use it:

• Python scripts that use the transformers library
• Vector databases and LLM tools such as Langchain, which the card says works with BGE models
• Windows, macOS and Linux with Python

Use case | Who | What they get back
Semantic search | Search developers | Vectors whose closeness reflects similar meaning
Clustering | Data analysts | Vectors that group similar texts together
Classification | Machine learning engineers | Features to train a small classifier on

How to set it up: You need these first:

• Python on Windows, macOS or Linux
• The libraries named in the card's usage section
• About 96 MB of disk space for the F32 weights

1. Install or update the Hugging Face download client.
2. Download the model files to your computer.
3. Follow the usage section of the model card to load the model.
4. Embed a few Chinese sentences and compare the vectors.

pip install -U huggingface_hub
hf download BAAI/bge-small-zh-v1.5

How to use it: You pass Chinese sentences to the model and get one vector for each. Closer vectors mean closer meaning. The text available here cuts off before the usage section, so follow the full card for code and for whether a query prefix helps.

Tool | What it does
FlagEmbedding | The BAAI project that holds this model's code, usage and fine tuning
bge-reranker models | Cross-encoders that re-rank the top results from an embedding model, more accurate but slower
C-MTEB | A Chinese benchmark of 31 test datasets for embedding models
Langchain | A framework the card says has BGE models built in
Query instruction | An optional prefix for retrieval queries, listed in the card's model table

Example: You embed three saved Chinese answers and one customer question. You compare the question's vector with each answer's vector and take the closest answer as the best match.

For newcomers and experts: An embedding turns a sentence into numbers so that similar meanings land close together. A good first try is embedding two related sentences and one unrelated sentence, then checking which pair is closest.

For readers who know the field, this is a dense embedding model, and the card says its reranker models are more accurate but less efficient. The card advises using a reranker to reorder the top results that an embedding model returns. Version 1.5 targets the similarity distribution issue of the first release.

Chinese text goes in
The model turns it into a vector
Vectors are compared or stored
Closest matches are returned
A reranker can reorder the top results

Good to know:

• Cut: the text available here ends inside the card's model table, before usage and evaluation.
• Version: the 1.5 models aim for a more reasonable similarity distribution.
• Reranking: the card recommends a reranker to reorder the top results from an embedding model.
• Language: Hugging Face lists Chinese.
• Size: 24M parameters, about 96 MB in F32.
• Versions: Hugging Face lists 11 quantized and 12 fine tuned versions made by others.
• License: MIT.
• Cost to run: none named. The weights are about 96 MB in F32.
• Data it reaches: the text you embed, on your own machine. The text available here names no network use.
• Not for: text in other languages, since the card lists separate English models.

On Hugging Face, 13 November 2026

• Task: feature extraction
• Library: transformers
• Parameters: 24M
• Weights: about 96 MB in F32; about 12 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,944,941
• Likes: 171
• License: mit
• Languages: zh
• Quantized versions: 11
• Fine-tunes: 12
• Tags: transformers, pytorch, safetensors, bert, feature-extraction, text-embeddings-inference

Links

• Model: https://huggingface.co/BAAI/bge-small-zh-v1.5
• Files: https://huggingface.co/BAAI/bge-small-zh-v1.5/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download BAAI/bge-small-zh-v1.5

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