---
license: apache-2.0
library_name: coreml
tags:
  - core-ml
  - siglip
  - sentence-transformers
  - embeddings
  - on-device
base_model:
  - google/siglip-base-patch16-384
  - sentence-transformers/all-MiniLM-L6-v2
---

# gallerywise/coreml-embeddings

Core ML (`.mlpackage`, fp16, `cpu_and_gpu`) conversions of the embedding backbones
used by the on-device pipeline in **gallerywise.ai**. Converted with
[`scripts/coreml/convert_embeddings.py`](https://github.com/ADR-007/aibom-macos)
(AIB-72 / AIB-119); loaded at runtime through pyobjc Core ML with **no torch** in
the shipped app.

## Contents

| File | Source model | Notes |
|------|--------------|-------|
| `siglip_vision.mlpackage.zip` | `google/siglip-base-patch16-384` | vision tower, 768-D |
| `siglip_text.mlpackage.zip` | `google/siglip-base-patch16-384` | text tower, 768-D |
| `text_embed.mlpackage.zip` | `sentence-transformers/all-MiniLM-L6-v2` | 384-D; mean-pool + L2-normalize baked into the graph (fixed seq-len 64) |
| `siglip_logit_params.json` | — | SigLIP logit scale/bias |

The `.mlpackage` bundles are directories, so they are hosted **zipped** (each
archive contains exactly one top-level `*.mlpackage/` entry). The client fetches,
sha256-verifies, and unzips them on first launch.

## License & attribution

Apache-2.0, inherited from both source models
([SigLIP](https://huggingface.co/google/siglip-base-patch16-384),
[all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2)),
which are themselves Apache-2.0. These are format conversions (Core ML) of those
weights — no retraining or fine-tuning.

## Intended use

The gallerywise.ai macOS app's embedding stages (semantic search, zero-shot tags,
near-duplicate / moment grouping). Not a general-purpose endpoint.