suhunhan95/sukurini-models

Sukurini Core ML models

Core ML exports used by Sukurini, a macOS screenshot manager, for on-device semantic image search. The app ships without these weights; they are downloaded once when the user turns on AI image search, verified against a pinned SHA-256, compiled to .mlmodelc, and run locally on the Apple Neural Engine. Nothing leaves the machine.

Models

vitb32-laion2b-fp16

Core ML FP16 export of OpenCLIP ViT-B/32 trained on LAION-2B (laion/CLIP-ViT-B-32-laion2B-s34B-b79K). Image and text encoders are exported separately so the text tower can be loaded only at query time.

File	Bytes	Purpose
image.mlpackage/Data/com.apple.CoreML/weights/weight.bin	175,712,384	Image encoder weights
image.mlpackage/Data/com.apple.CoreML/model.mlmodel	138,261	Image encoder graph
image.mlpackage/Manifest.json	617	Core ML package manifest
text.mlpackage/Data/com.apple.CoreML/weights/weight.bin	126,881,920	Text encoder weights
text.mlpackage/Data/com.apple.CoreML/model.mlmodel	137,570	Text encoder graph
text.mlpackage/Manifest.json	617	Core ML package manifest
tokenizer/merges.txt	524,605	CLIP byte-level BPE merges

Interface

• Image encoder: input image, 224x224 RGB; output final_emb_1, shape [1, 512], float32. Mean/std normalisation is baked into the graph, so the input is plain 0-255 RGB.
• Text encoder: input text, shape [1, 77] int32 CLIP BPE token ids; output final_emb_1, shape [1, 512], float32.
• Embeddings are not L2 normalised; normalise before computing cosine similarity.

Measured on an Apple M5 Pro, macOS 26.5.2, .cpuAndNeuralEngine

• Image encoder: 1.35 ms steady-state inference, 27.0 MB resident
• Text encoder: 0.97 ms steady-state inference, 19.5 MB resident

Licence and attribution

The weights are derived from OpenCLIP, released under the MIT Licence.

Copyright (c) 2012-2021 Gabriel Ilharco, Mitchell Wortsman, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, John Miller, Hongseok Namkoong, Hannaneh Hajishirzi, Ali Farhadi, Ludwig Schmidt.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Training data attribution: LAION-2B. See the base model card for details and intended-use guidance.