kanevry/siglip2-base-coreml

SigLIP 2 B/16-256 — Core ML (mirror)

Core ML build of SigLIP 2 (base, patch16, 256px) — 8-bit palettized weights, ready to run on-device on Apple Silicon. This repository is a bit-identical mirror re-hosted by AgenticCutter for distribution control (see Attribution).

Both encoders emit L2-normalized 768-dimensional embeddings, so an image embedding and a text embedding can be compared directly via cosine similarity (dot product) for zero-shot image-text retrieval and on-device visual search.

Attribution

• Base weights: google/siglip2-base-patch16-256, Google, licensed Apache-2.0.
• Core ML conversion: palmier-io/siglip2-base-coreml (8-bit palettized Core ML conversion of the SigLIP 2 base weights), licensed Apache-2.0. The three model files below are redistributed unmodified (same bytes, same sha256) from that source.
• This repository: a bit-identical, sha256-verified re-host of the palmier-io Core ML artifacts, maintained by AgenticCutter so the app can resolve a stable, first-party-controlled download URL. The model is loaded at runtime (on first use of on-device visual search) and is never bundled inside the AgenticCutter application binary.

Files

File	Description
ImageEncoder.mlpackage.zip	Core ML image encoder, zipped .mlpackage. Outputs a 768-d L2-normalized image embedding.
TextEncoder.mlpackage.zip	Core ML text encoder, zipped .mlpackage. Outputs a 768-d L2-normalized text embedding for the same joint space.
tokenizer.zip	Tokenizer assets required by the text encoder.
manifest.json	Per-file sha256 + byte-size manifest, plus model metadata (embedding dim, image size, context length).

Verifying integrity

Every file's sha256 is pinned in manifest.json and in AgenticCutter's own SearchIndexConfig.swift build manifest. Verify after download:

shasum -a 256 ImageEncoder.mlpackage.zip TextEncoder.mlpackage.zip tokenizer.zip

Expected hashes are listed in manifest.json in this repository.