---
license: mit
base_model: intfloat/multilingual-e5-small
tags:
  - coreml
  - sentence-similarity
  - multilingual
  - macos
library_name: coreml
---

# multilingual-e5-small — Core ML

Core ML conversion of [intfloat/multilingual-e5-small](https://huggingface.co/intfloat/multilingual-e5-small)
(revision `614241f622f53c4eeff9890bdc4f31cfecc418b3`), used by the Lenz macOS video editor to search
talk transcripts by meaning, on device. This repository only hosts the files the app downloads once.

## Files

| File | Contents |
|---|---|
| `SentenceEncoder.mlpackage.zip` | 12-layer encoder, int8 linear-symmetric weights; inputs `input_ids`, `attention_mask` (int32, batch 1–64 × tokens 1–512); output `hidden_states` (float32, batch × tokens × 384) |
| `tokenizer.zip` | XLM-RoBERTa tokenizer files (`tokenizer.json`, config) for swift-transformers |
| `manifest.json` | File names, sha256s, sizes, model dims |

## Usage notes

- Prefix every text with its role: `passage: ` for indexed sentences, `query: ` for searches.
  Same-prefix use measurably hurts retrieval.
- Sentence vector = mean of `hidden_states` over tokens where `attention_mask` is 1, then L2-normalised;
  similarity is a dot product.
- Pad with `<pad>` (id 1) and a zero mask; truncate to 512 tokens keeping the final `</s>`.
- Flexible shapes run fastest on the CPU compute unit (≈0.9 ms per sentence on an M2 Max).
- Conversion is parity-gated: pooled vectors match PyTorch at cosine ≥ 0.99 on EN/DE fixtures
  (int8 worst 0.99983). Source: `models/multilingual-e5-small/convert.py` in the Lenz repository.

## Versioning

Files in this repo are immutable once published. Re-conversions are published as new versions, never
overwrites.

## License

MIT, same as the original weights by intfloat (Liang Wang et al., "Multilingual E5 Text Embeddings",
arXiv:2402.05672). This repository redistributes a converted form of those weights without modification
to their values beyond int8 weight quantization.