README.md

---
base_model: mixedbread-ai/mxbai-edge-colbert-v0-32m
license: apache-2.0
library_name: coreml
pipeline_tag: sentence-similarity
tags:
  - coreml
  - colbert
  - multi-vector
  - late-interaction
  - feature-extraction
  - sentence-similarity
  - apple
  - on-device
---

# MXBAIEdgeColbert (Core ML)

A [Core ML](https://developer.apple.com/documentation/coreml) export of
[`mixedbread-ai/mxbai-edge-colbert-v0-32m`](https://huggingface.co/mixedbread-ai/mxbai-edge-colbert-v0-32m)
for on-device ColBERT late-interaction (multi-vector) encoding on Apple platforms.

## Base model

- **Base model:** [`mixedbread-ai/mxbai-edge-colbert-v0-32m`](https://huggingface.co/mixedbread-ai/mxbai-edge-colbert-v0-32m)

This repository contains only the compiled Core ML encoder (`MXBAIEdgeColbert.mlmodelc`).
All weights are derived from the base model above; no additional training was performed.

## What was exported

The exported program wraps the base transformer plus its three ColBERT `Dense`
projection layers (384 → 768 → 768 → 64, no bias / no activation), producing
per-token embeddings that are L2-normalized and masked by the attention mask.

| Property | Value |
| --- | --- |
| Inputs | `input_ids` and `attention_mask`, `int32`, static shape `(1, 256)` |
| Output | `token_embeddings` (per-token, dim 64, L2-normalized, padding zeroed) |
| Compute precision | `float16` |
| Minimum deployment target | iOS 18 / macOS 15 |

Tokenization and the ColBERT `[Q]` / `[D]` prefixing and skiplist masking are
performed by the host application; this model performs encoding only.

## Usage

Intended for use as the encoder in an on-device ColBERT/PLAID retrieval
pipeline. Load the compiled model with Core ML, feed padded `int32`
`input_ids` / `attention_mask`, and read `token_embeddings`.

## License

Released under the Apache-2.0 license, following the base model.