paraphrase-multilingual-MiniLM-L12-v2

paraphrase-multilingual-MiniLM-L12-v2

Turns sentences and paragraphs into 384-number vectors for semantic search, clustering and sentence similarity, with a multilingual model from the sentence-transformers project. For developers building search or matching.

Made by sentence-transformers 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: This model turns a sentence or a paragraph into a vector of 384 numbers. Texts with a similar meaning get vectors that sit close together, so you can use them for semantic search or for clustering. It is a multilingual BERT-based model with 118M parameters, trained by the sentence-transformers project.

Who it suits: It suits developers and analysts who need to compare texts by meaning instead of by exact words.

Industry | Who uses it | How they use it | Example
Search | Application developers | Match a typed question to documents by meaning | Finding help articles that answer a customer's question
Customer support | Support tool builders | Group messages that say the same thing in different words | Clustering tickets about one problem
Data analysis | Analysts | Cluster large piles of short text | Sorting survey answers into themes
Language technology | NLP researchers | Use a small encoder as a baseline for sentence tasks | Comparing it with another embedding model

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

Where to use it:

• Python programs on Windows, macOS or Linux that use the sentence-transformers library
• Python code that uses Hugging Face Transformers and its own pooling step
• Search, clustering and matching features inside an application

Use case | Who | What they get back
Semantic search | Developers | One 384-number vector per sentence, ready to compare
Clustering | Analysts | Vectors that put similar texts close together
Sentence similarity | Researchers | Embeddings to score how close two sentences are

How to set it up:

• Python
• The sentence-transformers library, or Transformers and torch for the plain route
• A connection to the Hugging Face hub the first time the model loads

1. Install the sentence-transformers library.
2. Load the model by its name in your Python code.
3. Encode a list of sentences.
4. Optionally, download the files to your own computer with the hub tool.

pip install -U sentence-transformers

pip install -U huggingface_hub
hf download sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2

How to use it: Load the model, pass a list of sentences to encode, and read back one embedding for each. With plain Transformers you tokenize the input, run the model, then average the token embeddings yourself, taking the attention mask into account.

from sentence_transformers import SentenceTransformer
sentences = ["This is an example sentence", "Each sentence is converted"]

model = SentenceTransformer('sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2')
embeddings = model.encode(sentences)
print(embeddings)

Tool | What it does
SentenceTransformer | Loads the model from its name
model.encode | Turns a list of sentences into embeddings
AutoTokenizer and AutoModel | The plain Transformers route, which needs a pooling step of your own
mean_pooling | Averages the token embeddings, using the attention mask for a correct average

Example: You pass two sentences, one saying this is an example and one saying each sentence is converted. Printing the result shows two embeddings, each a vector of 384 numbers.

For newcomers and experts: An embedding is a list of numbers that stands for the meaning of a text. Texts that mean alike get lists that are close to each other. A good first try is to install the library, run the snippet above and print the vectors.

Under the hood, the architecture is a BERT model with a maximum sequence length of 128, followed by pooling to 384 dimensions. The pooling averages tokens, with the class-token and max modes switched off. Longer inputs are cut at the limit. The hub's language list begins with Arabic, Bulgarian, Catalan, Czech, Danish, German, Greek and English.

Sentences go in
Tokenizer and BERT model
Mean pooling over the tokens
A 384-number vector comes out

Good to know:

Title: Weight size
F32: 471 MB
4-bit estimate: 59 MB

• The model reads at most 128 tokens of each input.
• The hub shows 46 quantized and 354 fine-tuned versions made by other people.
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: none named beyond the sentences you pass in, and the model files it fetches from the Hugging Face hub on first load
• Not for: none named, and it only produces vectors, so it does not write or answer text

On Hugging Face, 13 November 2026

• Task: sentence similarity
• Library: sentence-transformers
• Parameters: 118M
• Weights: about 471 MB in F32; about 59 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 26,540,125
• Likes: 1,505
• License: apache-2.0
• Languages: ar, bg, ca, cs, da, de, el, en
• Quantized versions: 48
• Fine-tunes: 354
• Tags: sentence-transformers, pytorch, onnx, safetensors, openvino, bert, feature-extraction, sentence-similarity, transformers, multilingual, text-embeddings-inference

Links

• Model: https://huggingface.co/sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
• Files: https://huggingface.co/sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2

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