| --- |
| license: apache-2.0 |
| base_model: google/siglip2-base-patch16-256 |
| tags: |
| - coreml |
| - siglip2 |
| - image-text-retrieval |
| library_name: coreml |
| --- |
| |
| # SigLIP2 Base — CoreML |
|
|
| CoreML build of [`google/siglip2-base-patch16-256`](https://huggingface.co/google/siglip2-base-patch16-256), |
| packaged for on-device visual + text search in [Cutroom](https://github.com/kuluruvineeth/cutroom). |
|
|
| Runs fully on-device via the Apple Neural Engine — no image or query ever leaves the machine. |
|
|
| ## Files |
|
|
| | File | Purpose | Bytes | |
| |---|---|---| |
| | `ImageEncoder.mlpackage.zip` | image tower → 768-d embedding | 91,700,398 | |
| | `TextEncoder.mlpackage.zip` | text tower → 768-d embedding | 258,593,083 | |
| | `tokenizer.zip` | tokenizer assets (64-token context) | 5,460,173 | |
|
|
| Each archive is SHA256-pinned by the client and verified after download. |
|
|
| ## Model |
|
|
| - Architecture: SigLIP2, base, patch16, 256×256 input |
| - Embedding dimension: 768 |
| - Context length: 64 tokens |
|
|
| Image and text are embedded into a shared space; retrieval is cosine similarity between them. |
|
|
| ## License |
|
|
| Apache 2.0, inherited from the base model. Copyright the SigLIP2 authors, Google LLC. |
| This repository redistributes a format conversion of those weights; the weights themselves are unmodified. |
|
|
