voyage-large-2 can handle multilingual text to some extent, but for multilingual retrieval specifically, Voyage positions a dedicated model—voyage-multilingual-2—as the optimized choice. Voyage’s announcement for voyage-multilingual-2 describes it as optimized for multilingual retrieval and RAG, and other ecosystem writeups describe it as supporting many languages (including Japanese and Korean) for multilingual retrieval scenarios. The Zilliz model comparison table also distinguishes voyage-multilingual-2 from voyage-large-2, explicitly labeling voyage-multilingual-2 as “optimized for multilingual retrieval and RAG.”
What this means for developers is: if you have mostly English content, or your multilingual needs are light (a handful of languages, limited cross-lingual queries), voyage-large-2 may still work acceptably depending on your corpus and evaluation results. But if your product needs true cross-lingual retrieval—like a Japanese query reliably retrieving English docs, or a Spanish query retrieving Japanese ticket summaries—you should treat multilingual performance as a first-class requirement and evaluate a model that’s explicitly optimized for it. The test is straightforward: create a multilingual relevance set (queries in language A, relevant passages in language B) and measure recall@k. If recall is inconsistent across languages, it’s a signal that you need a multilingual-optimized embedding model rather than assuming a generalist model will generalize.
No matter which model you choose, the way you operationalize multilingual retrieval is the same. You embed documents and queries into the same vector space, store vectors with metadata like lang, source, and tenant_id, and then decide whether you want cross-language search (no language filter) or same-language search (apply a lang == query_lang filter). A vector database such as Milvus or Zilliz Cloud is useful here because it lets you combine semantic search with language-aware filtering and partitioning—so you can offer toggles like “search all languages” vs “search Japanese only,” and keep latency stable even as the dataset grows. If multilingual effectiveness is central to your app, the most reliable path is to benchmark your own queries and content, then pick the Voyage model that matches that requirement explicitly.
For more information, click here: https://zilliz.com/ai-models/voyage-large-2