dayze/siglip2-base-patch16-256-coreml

SigLIP 2 B/16-256 for Core ML (Dayze)

Core ML conversion of google/siglip2-base-patch16-256 (revision 3f9f96cb90da5dbc758b01813f2f6f1aee24c1ab), used by Dayze Desktop for on-device photo search. Photos never leave the device; this repository only hosts the model files.

File	Tower	Compression	Size
siglip2_b16_256_Image_pal8.mlpackage	image	8-bit palettized	~89 MB
siglip2_b16_256_Text_pal8ng.mlpackage	text	8-bit palettized (embedding not grouped)	~458 MB

Tokenizer and preprocessing files are copied unchanged from the base model. SHA256SUMS lists a checksum for every file; clients pin a repository revision and verify these hashes.

Run with MLComputeUnits.cpuAndNeuralEngine. Measured on an M5 Mac: ~5.7 ms per image, ~1.5 ms per text query. COCO-1k text→image R@1 71.4 (Core ML) vs 58.1 for TinyCLIP-39M.

Licence and attribution

The model weights and code are by Google, released under the Apache License 2.0 (model card, big_vision). Other SigLIP 2 materials are under CC-BY 4.0.

Modifications by Dayze: conversion to Core ML and 8-bit weight palettization. No retraining.

Citation: Tschannen et al., SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features, 2025. arXiv:2502.14786.