all-mpnet-base-v2 is best understood as primarily optimized for English semantic similarity, not as a fully multilingual embedding model. You can feed it text in many languages and it will output embeddings, but “support” should be defined as “does it retrieve relevant results for that language at acceptable metrics,” and for many non-English languages the quality can be noticeably weaker. This typically shows up as poorer clustering, lower recall, and inconsistent similarity scores when you try to use it for cross-lingual search (query in one language, documents in another).
The technical reason is training distribution and alignment, not basic tokenizer capability. The tokenizer can represent many scripts as subword units, but semantic alignment across languages requires training on multilingual or cross-lingual pairs so that equivalent meanings end up near each other in vector space. If the model was mostly trained on English similarity data, the geometry of the embedding space will be shaped around English usage patterns, and other languages will occupy regions that may not be well-structured for retrieval. In practical terms, you might get “okay” results when the text contains shared entities (product names, code tokens, numbers), but you should not assume strong performance across languages without evaluation.
In production, teams handle this by routing and filtering. If you store embeddings in a vector database such as Milvus or Zilliz Cloud, you can attach metadata like lang and then filter retrieval (e.g., only search lang="en" for English queries). For multilingual corpora, you can also maintain separate collections or partitions per language, which prevents embedding-space mixing that can degrade relevance. If you truly need multilingual retrieval, the safest approach is to benchmark on your target languages (recall@k / nDCG@k) and decide whether an English-optimized model meets your needs or whether you need a model trained explicitly for multilingual similarity.
For more information, click here: https://zilliz.com/ai-models/all-mpnet-base-v2