BGE-M3: embeddings that truly speak French

#The French problem

An embedding model learns its notion of similarity from its training corpus. If that corpus is 90% English, the model correctly positions English texts relative to one another, but handles French texts much less precisely. In practice, on a French document corpus, this means relevant passages fail to surface while off-topic passages do—with no error message, no warning in the logs, and often without anyone noticing until a user reports an irrelevant answer.

That's why the answer to “which embedding model?” differs by language. Public benchmarks (such as MTEB) are heavily dominated by English-language tasks; they don't predict what will happen on your documents or specialized vocabulary. That's exactly the use case for which the Chinese lab BAAI (Beijing Academy of Artificial Intelligence) designed BGE-M3: the name stands for Multi-Functionality, Multi-Linguality, Multi-Granularity—three search functions, more than 100 languages, and inputs ranging from short sentences to long documents, combined in a single model rather than spread across multiple tools.

#What BGE-M3 brings

Multilingual by design

According to its official profile, the model supports more than 100 working languages; it handles French correctly and also lets you query an English corpus in French.

Long context: 8 192 tokens

It accepts significantly longer passages than most embedding models (often limited to 512 tokens), giving you more flexibility in the size of indexed chunks.

Three representations at once

Dense, lexical, and multi-vector, produced in a single pass with no additional compute cost, according to BAAI.

MIT License

Usable in business without particular restrictions; always check the model card when integrating it, as the license may change from one version to another.

No instruction to add

Unlike older BGE models, BGE-M3 no longer requires adding an instruction to queries: use it directly, which simplifies integration.

#Dense, lexical, multi-vector: what it is for

The benefit is being able to run hybrid search without running two separate models: the official page explicitly recommends the hybrid search + reranking combination, using lexical weights obtained “at no additional cost” alongside the dense embedding. The lexical component precisely addresses the dense component's failures, while the multivector component reranks the twenty or so selected candidates. Not all vector databases can use all three outputs, but dense-only is already enough for typical use—it's also the simplest configuration to implement when getting started, before adding the lexical layer if accuracy is still lacking.

#Use it locally

The fastest way to test BGE-M3 locally is Ollama, which distributes a quantized version of the model under the name bge-m3:567m — 567 million parameters, a download of approximately 1.2 GB, and an advertised context window of 8,000 tokens on the model page. It is built on XLM-RoBERTa-large, whose context length was first extended to 8,192 tokens through dedicated pretraining (RetroMAE), followed by unified fine-tuning for the three retrieval tasks in a single training stage, rather than three separate models to maintain.

ollama pull bge-m3
ollama run bge-m3 "Quel est le délai de rétractation légal en France ?"

For a complete RAG pipeline, only dense output is directly usable by most standard vector databases (Qdrant, Milvus, pgvector) without specific configuration; lexical and multi-vector outputs require an engine that knows how to combine them, as documented, for example, by the Milvus and Vespa integrations cited by BAAI.

In a standard RAG pipeline with a local LLM, BGE-M3 simply replaces the default embedding model: each document chunk is encoded once during indexing, the user’s question is encoded for each query, and the best retrieved passages are then sent to the LLM (Qwen, Mistral, Llama…) loaded under Ollama or LM Studio. Choosing the embedding model and choosing the generation model are two independent decisions: nothing requires using models from the same family for both steps, and in practice that is rarely the case.

#Check against your own documents

A product page's “more than 100 languages” is not a guarantee of uniform quality: it is coverage, not a per-language score. BAAI evaluates BGE-M3 on MIRACL (multilingual retrieval) and MLDR, a long-document retrieval dataset covering 13 languages that the lab published specifically for this model, with the corresponding evaluation pipeline openly available. These evaluations provide a general trend measured on standardized research corpora; they do not replace testing on your own corpus, in your own domain, with your own technical vocabulary and your own question phrasing.

1. 01
   Create a small representative sample

   Twenty to thirty real documents from the target corpus, with their usual diversity of length and vocabulary—not a handpicked selection.
2. 02
   Write questions the way your users would ask them

   Ten to twenty real questions, including questions that use synonyms or wording different from the source document.
3. 03
   Compare two or three models on the same sample

   BGE-M3 against a well-regarded English-language model and, if possible, another multilingual model. Note how many times the correct passage appears in the top three results.
4. 04
   Decide before the bulk import

   The test takes half a day; changing models after indexing an entire corpus means re-encoding everything, as noted above.

#What it costs

This is a heavier embedding model than the small English-language models currently in fashion: its 567 million parameters and 1.2 GB download make it mid-weight for the category, far from the models with a few tens of millions of parameters that dominate pure speed rankings. It encodes more slowly and uses more memory. You’ll notice this during the initial indexing of a large corpus; for a single question, the difference is imperceptible compared with the generation time of the language model that follows.

Another consequence: its dense vectors have a dimensionality of 1,024, so the index takes up more space than models with 384 or 768 dimensions. With one million passages, the arithmetic is no longer negligible, and it is time to look at your vector database’s compression options.

The contrast is clear: at the same dimension (1,024), BGE-M3 accepts sixteen times more tokens per passage than bge-large-en-v1.5 and covers more than 100 languages, whereas BAAI's entire “en” family is limited to English. That's the price, in download size and encoding time, of not having to choose a different model for each corpus language or maintain multiple separate indexes based on the language of incoming documents.

#When to choose it and when to skip it

• The landscape of embedding models for French
• Run an embedding model locally
• Re-rank candidates for greater accuracy
• Store these vectors in PostgreSQL with pgvector
• Source: official BGE-M3 documentation on Hugging Face
• Source: bge-m3 in the Ollama library
• Source: official BGE-M3 repository and code (FlagEmbedding)

#FAQ