hackerker-roylo/siglip2-base-image-8bit-coreml

SigLIP2 base (patch16, 256) image encoder — Core ML, 8-bit palettized

This is the image encoder only of Google's SigLIP2 (google/siglip2-base-patch16-256), as a Core ML .mlpackage with weights palettized to 8 bits (~93 MB, down from ~185 MB fp16).

It is hosted for the Life in Photos iOS app, which downloads it on Wi-Fi and uses it to recognize "reminder" photos (receipts, signs, screenshots, documents) so they are kept out of photo moments. This is a pre-production hosting location.

Credits

• Model: SigLIP2 by Google — google/siglip2-base-patch16-256, Apache-2.0.
• Core ML conversion: FluidInference/siglip2-base-patch16-256-coreml, Apache-2.0.
• This variant: FluidInference's fp16 image encoder palettized to 8 bits with coremltools (OpPalettizerConfig(nbits=8)). No retraining. Image embeddings stay within ~0.997 cosine similarity of the fp16 model.

Files

• siglip2-base-patch16-256-image-pal8.mlpackage/ — the Core ML package (input: 256×256 image).
• LICENSE — Apache License 2.0.

License

Apache License 2.0, same as the upstream model and conversion. See LICENSE. Modified from the originals by 8-bit weight palettization.