sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 · Hugging Face
The `paraphrase-multilingual-MiniLM-L12-v2` model is a sentence-transformers model that maps sentences and paragraphs to a 384-dimensional dense vector space for tasks like clustering, semantic search, and paraphrase identification. It is based on Microsoft's MiniLM-L12-H384-uncased architecture and fine-tuned on a large multilingual paraphrase corpus, supporting over 50 languages. The model is released under Apache 2.0 and is available in PyTorch, TensorFlow, ONNX, OpenVINO, and SafeTensors formats.
- Depth
- 870
- Categories
- 1
- Index status
- Live
word-level signal
topic cluster links
Jun 26, 2026
Deep Brief
Overview and use cases
Overview
The paraphrase-multilingual-MiniLM-L12-v2 model is a sentence-transformers model designed to produce semantically meaningful sentence embeddings. It maps sentences and paragraphs to a 384-dimensional dense vector space, enabling tasks such as clustering, semantic search, and paraphrase identification. The model is based on the Microsoft MiniLM-L12-H384-uncased architecture and has been fine-tuned on a large corpus of paraphrases across multiple languages. It supports over 50 languages, making it suitable for multilingual applications. The model is hosted on Hugging Face at sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 and has garnered over 50 million downloads, reflecting its widespread adoption.
Capabilities
- Multilingual Sentence Embeddings: Generates embeddings for sentences in 50+ languages, allowing cross-lingual semantic similarity comparisons. Languages include Arabic, Bulgarian, Catalan, Czech, Danish, German, Greek, English, Spanish, Estonian, Persian, Finnish, French, Galician, Gujarati, Hebrew, Hindi, Croatian, Hungarian, Armenian, Indonesian, Italian, Japanese, Georgian, Korean, Kurdish, Lithuanian, Latvian, Macedonian, Mongolian, Marathi, Malay, Burmese, Norwegian Bokmål, Dutch, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Albanian, Serbian, Swedish, Thai, Turkish, Ukrainian, Urdu, Vietnamese, and more.
- Paraphrase Detection: Identifies whether two sentences are paraphrases of each other, useful for duplicate detection.
- Semantic Search: Enables retrieval of semantically similar sentences from a corpus, supporting both monolingual and cross-lingual search.
- Clustering: Groups similar sentences based on embedding similarity, facilitating document organization.
- Efficient Inference: Optimized for speed with a compact 12-layer transformer (MiniLM architecture), making it suitable for resource-constrained environments.
- Multiple Framework Support: Available in PyTorch, TensorFlow, ONNX, OpenVINO, and SafeTensors formats, providing flexibility for deployment.
- Sentence Similarity Pipeline: The model is tagged with
sentence-similaritypipeline, and can be used with Hugging Face Inference Endpoints.
Use cases
- Multilingual Semantic Search: Build search engines that understand queries in multiple languages and retrieve relevant documents regardless of language. For example, a query in English can retrieve documents in French or German.
- Paraphrase Mining: Detect duplicate or near-duplicate content in large text corpora, useful for plagiarism detection or content deduplication in multilingual datasets.
- Customer Support: Automatically route tickets to appropriate teams by comparing user queries to known issues across languages.
- Chatbots: Improve intent recognition by comparing user inputs to predefined intents across languages, enabling multilingual chatbots.
- Document Clustering: Organize large collections of documents into thematic groups based on semantic similarity, even when documents are in different languages.
- Cross-lingual Information Retrieval: Retrieve information in one language using queries in another, facilitating knowledge discovery across language barriers.
License & deployment
The model is released under the Apache 2.0 license, as indicated by the license:apache-2.0 tag on Hugging Face. It can be deployed using the sentence-transformers library, Hugging Face Transformers, or via ONNX Runtime for optimized inference. The model is also available on Hugging Face Hub and can be used with the sentence-transformers pipeline: sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2. For deployment, consider using containerized solutions like Docker or serverless platforms such as AWS Lambda or Hugging Face Inference Endpoints. The model is compatible with text-embeddings-inference and is deployable on Azure (tagged deploy:azure). Hardware considerations: The model is relatively lightweight (384-dimensional embeddings, 12 layers) and can run on CPU, but GPU acceleration is recommended for high-throughput scenarios. The model is also available in ONNX and OpenVINO formats for edge deployment.
Alternatives
- paraphrase-multilingual-MiniLM-L12-v1: An earlier version of the same model, with slightly different training data and performance.
- paraphrase-MiniLM-L6-v2: A smaller, faster English-only version with 6 layers and 384-dimensional embeddings.
- distiluse-base-multilingual-cased-v2: A multilingual sentence embedding model based on DistilBERT, supporting 50+ languages with 512-dimensional embeddings.
- LaBSE: A multilingual BERT-based model for sentence embeddings, supporting 109 languages, but larger and slower.
- all-MiniLM-L12-v2: A general-purpose English sentence embedding model with the same architecture but trained on diverse data.
- multilingual-e5-large: A larger multilingual embedding model with higher accuracy but more resource-intensive.
FAQ
Q: What is the dimensionality of the embeddings? A: The model produces 384-dimensional dense vectors.
Q: How many languages does the model support? A: Over 50 languages, including Arabic, Chinese, English, French, German, Spanish, and many more. A full list is available on the Hugging Face model card.
Q: What is the architecture of the model? A: It is based on Microsoft's MiniLM-L12-H384-uncased, a 12-layer transformer with 384 hidden dimensions.
Q: Can I use this model for cross-lingual semantic search? A: Yes, the model is designed for cross-lingual applications and can compare sentences across different languages.
Q: What license is the model released under? A: Apache 2.0.
Q: How can I deploy this model?
A: You can use the sentence-transformers library, Hugging Face Transformers, ONNX Runtime, or Hugging Face Inference Endpoints. It is also compatible with text-embeddings-inference and Azure deployment.
Q: Is there a smaller version available?
A: Yes, the paraphrase-MiniLM-L6-v2 is a smaller English-only version with 6 layers.
Q: What is the difference between v1 and v2 of this model? A: v2 has been fine-tuned on additional data and may show improved performance on certain tasks. Specific details are available in the model card.
Q: Can I use this model for clustering? A: Yes, the embeddings can be used with clustering algorithms like K-means or DBSCAN.
Q: Does the model support ONNX? A: Yes, ONNX format is available for optimized inference.
Keep Exploring
Related AI models
lpiccinelli
lpiccinelli/unidepth-v2-vitl14 · Hugging Face
UniDepth v2 (ViT-L/14) is a 0.4B-parameter PyTorch model for monocular metric depth estimation, leveraging a Vision Transformer backbone to predict scale-aware depth maps from single RGB images.
Qwen
Qwen/Qwen3-8B · Hugging Face
Qwen/Qwen3-8B is an 8-billion parameter large language model developed by Alibaba Cloud's Qwen team, available on Hugging Face under the Apache 2.0 license. It is designed for advanced text generation, conversational AI, and complex reasoning tasks, featuring support for tool use and long-context understanding.
google-t5
google-t5/t5-small · Hugging Face
T5-small is a compact, multilingual encoder-decoder transformer model developed by Google and hosted on Hugging Face. Licensed under Apache 2.0, it supports text-to-text generation tasks such as translation, summarization, and question answering across multiple languages including English, French, German, and Romanian. It is optimized for low-latency inference and efficient deployment on various hardware backends.
Qwen
Qwen/Qwen3-0.6B · Hugging Face
Qwen/Qwen3-0.6B is a lightweight, open-source large language model developed by Alibaba Cloud's Qwen team. Released under the permissive Apache 2.0 license, this 0.6-billion-parameter model is designed for high-efficiency deployment on consumer-grade hardware, including CPUs and low-end GPUs. It features a 32,768-token context window, multilingual capabilities, and native support for tool-use and function calling, making it an ideal candidate for edge computing, real-time chatbots, and resource-constrained environments.
Ollama
llava
LLaVA (Large Language and Vision Assistant) is an open-source, end-to-end trained multimodal model that combines a vision encoder with a large language model (Vicuna) to enable visual understanding and reasoning. Developed by researchers from Microsoft, the University of Wisconsin-Madison, and Columbia University, LLaVA allows users to interact with images through natural language, supporting tasks such as visual question answering, image captioning, and complex visual reasoning. The latest iteration, LLaVA 1.6, introduces significant improvements in input resolution, visual instruction tuning, and logical reasoning capabilities. It is widely accessible for local deployment via Ollama, offering privacy, offline functionality, and flexible hardware requirements across various model sizes.
BAAI
BAAI/bge-small-en-v1.5 · Hugging Face
BAAI/bge-small-en-v1.5 is a lightweight, high-performance sentence embedding model developed by the Beijing Academy of Artificial Intelligence (BAAI). Based on the BERT architecture with approximately 33 million parameters, it generates 384-dimensional vector representations optimized for English text. It is widely used for semantic search, retrieval-augmented generation (RAG), and clustering tasks, offering a balance between computational efficiency and accuracy under the permissive MIT License.
Site Discovery
Keep exploring ToolSeekAI
Move from model intelligence into tools, news, and rankings for a stronger AI decision path.