multilingual-e5-small — GGUF (F16)

A GGUF (F16) conversion of intfloat/multilingual-e5-small, for serving embeddings with llama.cpp / llama-server --embeddings.

• Dimensions: 384
• Pooling: mean (required)
• Normalization: L2 (use cosine similarity)
• Context length: 512 tokens
• Architecture: XLM-RoBERTa (BERT-family encoder)
• License: MIT (inherited from the base model)

Prefixes are mandatory

multilingual-e5-small was trained with asymmetric prefixes. You must prepend:

• query: to search queries
• passage: to documents being indexed

Retrieval quality collapses without them.

Usage with llama-server

llama-server -m multilingual-e5-small-f16.gguf --embeddings --pooling mean -c 512

Then POST to the OpenAI-compatible endpoint:

curl http://127.0.0.1:8080/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{"model": "multilingual-e5-small-f16.gguf", "input": ["query: what is retrieval augmented generation"]}'

Fidelity

F16 embeddings track the reference sentence-transformers output closely (cosine similarity min 0.9976, mean 0.9992 across a 20-sentence multilingual sample spanning English, French and Chinese); the small residual is dominated by SentencePiece tokenization differences, and top-1 ranking is preserved.

Conversion

Converted from the original safetensors weights with convert_hf_to_gguf.py from llama.cpp.