Multilingual Embedding (small, 141M) - bge-m3 compatible

A multilingual sentence and passage embedding model distilled from BAAI/bge-m3 into mmBERT-small. It outputs 1024-dimensional vectors in bge-m3's dense vector space, so you can:

• use it on its own for search, RAG, clustering and similarity in about 90 languages, 4x smaller than bge-m3;
• or query an existing bge-m3 index with it: embed queries with this model and keep the passages you already embedded with bge-m3 (see the "→ bge-m3 index" row below). That gives cheap query encoding on CPU or in the browser.

A larger, more accurate version is available as Horizon-Labs/multilingual-embedding-base. Apache-2.0. ONNX files for CPU and the browser (transformers.js) are included.

• Vectors are L2-normalised. Use cosine similarity (= dot product). No query/passage prefixes are needed.
• onnx/model_quantized.onnx (int8 embeddings, 269 MB; cosine >= 0.99997 to the fp32 model on test sentences) outputs the final normalised embedding as sentence_embedding.

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Horizon-Labs/multilingual-embedding-small")
q = model.encode(["How tall is the Eiffel Tower?"])
p = model.encode(["La tour Eiffel mesure 330 mètres.", "The Statue of Liberty is 93 metres tall."])
print(q @ p.T)   # higher = more similar

# Querying an existing bge-m3 index: passages embedded earlier with BAAI/bge-m3 (dense, normalised)
# scores = model.encode(queries) @ bge_m3_passage_vectors.T

transformers.js:

import { AutoTokenizer, AutoModel } from "@huggingface/transformers";
const tok = await AutoTokenizer.from_pretrained("Horizon-Labs/multilingual-embedding-small");
const model = await AutoModel.from_pretrained("Horizon-Labs/multilingual-embedding-small", { dtype: "q8" });
const { sentence_embedding } = await model(await tok(["Wie hoch ist der Eiffelturm?"], { padding: true, truncation: true }));

Evaluation

nDCG@10 for ranking each query's candidates by cosine similarity, on public reranking benchmarks (MTEB versions; used only for evaluation). MIRACL: 60 queries per language with 100 candidates each. Wikipedia: 60 queries per language with 9 candidates each. "Other" = ESCI (es, jp, us), RuBQ, T2Reranking, mMARCO-ja and AskUbuntu. Same script for every model (code/); multilingual-e5 is run with its "query: "/"passage: " prefixes.

• The table shows the released checkpoint. Means over two training seeds: small own .817 / against a bge-m3 index .818; base own .822 / against a bge-m3 index .816.
• Models ahead of this one (own embeddings): MIRACL: bge-m3; Wikipedia: none; other: bge-m3, multilingual-e5-small, multilingual-e5-base.
• This is a reranking-style test over given candidate lists, not a full-corpus retrieval benchmark. Rerankers on the same test: leaderboard.

Training

• Distillation: the student (mmBERT-small + mean pooling + a linear layer to 1024 dimensions) learns to reproduce bge-m3's normalised dense embeddings: cosine + squared-error loss, plus a loss on the in-batch similarity matrix that keeps the relative geometry.
• Text: about 4.65 million texts: snippets of 1-8 sentences and short spans from FineWeb-2 and FineWeb (ODC-BY) in 92 languages, plus about 850k search queries written by Qwen3.8-27B (Apache-2.0) for web passages in 47 languages.
• Code: code/ in this repository.

Limitations

• The student approximates bge-m3 and is weaker than it. The gap is larger for long passages, rare languages and specialised domains.
• Only bge-m3's dense vectors are reproduced, not its sparse or multi-vector (ColBERT) outputs.
• Max length 512 tokens (trained with 256). Split long documents into passages.