cstr/multilingual-e5-small-GGUF

multilingual-e5-small GGUF

GGUF format of intfloat/multilingual-e5-small for use with CrispEmbed and Ollama.

Files

File	Quantization	Size	cos vs HF
multilingual-e5-small.gguf	F32	455 MiB	1.0000 (reference)
multilingual-e5-small-q8_0.gguf	Q8_0	126 MiB	0.9999
multilingual-e5-small-q4_k.gguf	Q4_K	115 MiB	0.990
multilingual-e5-small-q4_k-imatrix.gguf	Q4_K + imatrix	115 MiB	not measured
multilingual-e5-small-iq4_xs.gguf	IQ4_XS	115 MiB	not measured

Sizes are MiB (what the file browser above reports). Auxiliary files: multilingual-e5-small.imatrix (131 KiB, the importance matrix used for the imatrix/IQ quants) and multilingual-e5-small-imatrix-ab.txt (the A/B notes from producing it).

Recommended: Q8_0. It is within 0.0001 cosine of the F32 reference while being 3.6× smaller, so F32 buys nothing you can measure. Drop to a 4-bit quant only when you are memory-bound: they save a further 11 MiB — about 9% — and Q4_K gives up an order of magnitude more accuracy (0.990) to do it. At this model size the 4-bit tier is rarely the right trade; it matters much more on the larger models in this registry.

Quick Start

CrispEmbed

./crispembed -m multilingual-e5-small "Hello world"
./crispembed-server -m multilingual-e5-small --port 8080

Ollama (with CrispStrobe fork)

# Create model
echo "FROM multilingual-e5-small-q8_0.gguf" > Modelfile
ollama create multilingual-e5-small -f Modelfile

# Embed
curl http://localhost:11434/api/embed -d '{"model":"multilingual-e5-small","input":["Hello world"]}'

Python (CrispEmbed)

from crispembed import CrispEmbed
model = CrispEmbed("multilingual-e5-small-q8_0.gguf")
vectors = model.encode(["Hello world", "Goodbye world"])

Model Details

Property	Value
Architecture	BERT
Parameters	118M
Embedding Dimension	384
Layers	12
Pooling	mean
Tokenizer	SentencePiece
Language	multilingual
Q8_0 vs HuggingFace	0.9999
Q4_K vs HuggingFace	0.990

Server API

CrispEmbed server supports four API dialects:

• POST /embed — native
• POST /v1/embeddings — OpenAI-compatible
• POST /api/embed — Ollama-compatible
• POST /api/embeddings — Ollama legacy

Credits

• Original model: intfloat/multilingual-e5-small
• Inference: CrispEmbed (MIT, ggml-based)

Provenance and EU AI Act Art. 53 note

• Upstream model: intfloat/multilingual-e5-small — published by intfloat.
• Upstream licence: mit. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.