mickael-palma/search-models

On-device search models for Apple silicon

Ready-to-load conversions of text, image and speech models for macOS on Apple silicon, gathered in one repo so they can be fetched with plain curl, with no Python, Xcode or conversion step on the Mac that uses them.

Nothing here is fine-tuned: every model is a format conversion or a copy of the upstream weights, listed below with its source and licence.

Path	What	Source	Licence
qwen3-embedding-0.6b-coreai.aimodel, qwen3-embedding-0.6b-coreai-tokenizer.bin	Text embedding, exported to Core AI (fp16, input 2–128 tokens, unpadded, output [1, 1024]), and a tokenizer table	Qwen/Qwen3-Embedding-0.6B @ 97b0c614	Apache 2.0
siglip2-base-patch16-224-{image,text}.mlpackage, siglip2-base-patch16-224-tokenizer.bin	Image–text embedding: the image and text towers as Core ML, and a text tokenizer table	google/siglip2-base-patch16-224 @ 75de2d55	Apache 2.0
bge-reranker-v2-m3-mlx/	Cross-encoder reranker, MLX	flaglow/BAAI-bge-reranker-v2-m3-mlx-fp16 @ a8ecc6ca, from BAAI/bge-reranker-v2-m3	Apache 2.0
whisper-large-v3-turbo-coreml/	Speech to text, Core ML compiled for the Neural Engine (632 MB palettized)	argmaxinc/whisperkit-coreml @ 0f63a780, openai_whisper-large-v3-v20240930_turbo_632MB; tokenizer from openai/whisper-large-v3 @ 06f233fe	MIT

The .bin tokenizer tables are a compact custom format derived from the upstream tokenizer.json; use the upstream tokenizer with any other runtime.

The Core AI model needs macOS 27.