---
license: mit
base_model: intfloat/multilingual-e5-base
library_name: gguf
tags:
  - feature-extraction
  - sentence-similarity
  - embeddings
  - gguf
pipeline_tag: feature-extraction
---

# multilingual-e5-base — GGUF (Q8_0)

GGUF build of [intfloat/multilingual-e5-base](https://huggingface.co/intfloat/multilingual-e5-base) — a multilingual XLM-RoBERTa embedding model (278M params, 768 dimensions, 100 languages, `passage:`/`query:` asymmetric prefixes).

Built **from first-party intfloat safetensors**, because no reputable quantizer publishes this model — this repo exists so downstream projects (vectorlink, TerminusDB) have an auditable supply chain rather than pulling a random community upload.

## Provenance

| Step | Tool | Pin |
|------|------|-----|
| Source weights | `intfloat/multilingual-e5-base` | commit `d128750597153bb5987e10b1c3493a34e5a4502a` |
| Conversion | llama.cpp `convert_hf_to_gguf.py` | release `b6987` |
| Quantization | `llama-quantize` → Q8_0 | llama.cpp `0.6.0` |

Build script: [`make-multilingual-e5-base-gguf.sh`](https://github.com/terminusdb-org/vectorlink/blob/main/scripts/make-multilingual-e5-base-gguf.sh) — downloads pinned first-party weights, verifies the safetensors SHA-256, converts, asserts mean pooling + 768-dim in the GGUF metadata, then quantizes.

## Verification

Same probe text, cosine similarity vs the original safetensors weights (f16 conversion): `0.9998`. Mean pooling, `passage:`/`query:` prefixes applied by the caller per E5 convention.

## Usage (Ollama)

```bash
ollama pull hf.co/vectorlink/multilingual-e5-base-GGUF:Q8_0
ollama cp hf.co/vectorlink/multilingual-e5-base-GGUF:Q8_0 multilingual-e5-base
curl localhost:11434/api/embed -d '{"model":"multilingual-e5-base","input":"query: test"}'  # → 768-dim
```

## License

MIT — inherited from `intfloat/multilingual-e5-base`. This repo performs format conversion and quantization only; no training, fine-tuning, or modification of weights. See `LICENSE`.