bge-m3

BGE‑M3 is an embedding model from the Beijing Academy of Artificial Intelligence for advanced retrieval. It supports three modes: dense, multi-vector and sparse. Its multigranular approach handles texts of varying length in more than 100 languages. It is optimized for semantic search, document ranking and RAG pipelines, balancing accuracy and inference speed.

• PROVIDER LOCATION: Non-EU
• PARAMETERS: ~567M
• MAX CONTEXT: 8.192 token
• LICENSE: MIT
• FEATURES: Dense retrieval, Sparse retrieval, Multi-vector (ColBERT)
• OUTPUT FORMATS: Embedding_vector

• Semantic search on long documents
• Multilingual RAG pipelines
• Result ranking and re-ranking
• Enterprise indexing and retrieval

curl --request POST \
--url https://apiaruba.it/aiplatform/v1/embeddings \
--header 'content-type: application/json' \
--header 'x-aruba-api-key: *****' \
--data '{
"model": "bge-m3",
"input": [
"primo testo",
"secondo testo"
]
}'

{
"model": "bge-m3",
"data": [
{
"embedding": [
0.01378626748919487,
0.031720615923404694,
-0.02318044938147068,
-0.007808151189237833,
0.0043615843169391155,
...
]
}
]
}

• multilingual‑e5‑large

• Hugging Face
• Paper