all-MiniLM-L12-v2 is best suited for English, but it is not strictly “English-only” in the sense that it will refuse other languages. You can input text in many languages and the model will still produce embeddings. However, the quality of those embeddings depends on how well the language was represented in the model’s training data, and for this model, English dominates. As a result, semantic similarity works best for English text and English-only retrieval tasks.
From a technical standpoint, this limitation comes from training distribution rather than architecture. The tokenizer can handle Unicode text and split words from many languages into subword units, but semantic alignment across languages requires explicit multilingual or cross-lingual training. If the model has not seen enough paired examples in a given language, the resulting vector space may not cluster meaningfully for that language. In practice, this means you might see acceptable results for languages that share vocabulary or structure with English, but weaker performance for others, especially in cross-lingual search scenarios.
In production systems, developers often address this by routing queries and documents by language. If you are using all-MiniLM-L12-v2, a common pattern is to detect language first and only apply the model to English content, while handling other languages separately. A vector database such as Milvus or Zilliz Cloud makes this easy by supporting metadata filters like lang="en". This ensures English queries only search English vectors, which improves relevance and avoids mixing poorly aligned embeddings from different languages.
For more information, click here: https://zilliz.com/ai-models/all-minilm-l12-v2