heypatkay/solume-sam3-coreml

SAM 3, Core ML (fp16), for Solume

A Core ML conversion of SAM 3 (Segment Anything with Concepts, Meta AI), the image model's vision trunk, text tower and detector, made for the text masks of Solume, a macOS photo app: describe what to select ("sky", "the red car", "all the lanterns") and SAM 3 finds every instance. Converted to Core ML by Solume. Solume downloads the files in sam3/ from a pinned commit of this repository and installs them only if every file matches the size and SHA-256 pinned in the app.

This repository holds only the converted model, its licence and this card. It holds no photos or user data.

Licence and use (please read)

SAM 3 is licensed under Meta's SAM License (last updated November 19, 2025). The full text, exactly as shipped in facebook/sam3, is in LICENSE, and a second copy travels inside sam3/ with the weights. These files are a derivative work of the SAM Materials and are distributed under the terms of that licence. Any use of these files, or of anything made from them, is subject to the SAM License, including its terms on trade controls (section 1(b)(iii) and (v): no use for activities subject to ITAR or end uses prohibited by trade controls, such as military or warfare purposes, nuclear industries or applications, espionage, or the development or use of guns or illegal weapons), on not reverse engineering or encouraging others to (1(b)(iv)), and on acknowledging the use of SAM Materials in published research (1(b)(ii)). By using or distributing any part of these files you agree to be bound by the licence. If you distribute them or a derivative, you must do so under the SAM License and include a copy of it.

Provenance

• Upstream weights: facebook/sam3 at revision 3c879f39826c281e95690f02c7821c4de09afae7, model.safetensors (3,439,938,512 bytes, SHA-256 6d06f0a5f84e435071fe6603e61d0b4cc7b40e0d39d487cfd4d67d8cc11cc14a), with its tokenizer files (vocab.json, merges.txt, copied unchanged) and LICENSE (7,352 bytes, SHA-256 bec48f70bd37bf8280a9d1ebf01642d26086f6122ba735baa08fe03c5a6e7448, copied unchanged).
• Conversion: Solume's scripts/sam3/convert.py sam3: transformers' Sam3Model (eager attention), exported with torch.export and converted with coremltools to three fp16 ML Programs. Transformers' attention-mask builders are replaced by plain additive masks and the ViT's position embeddings are tiled once so the graph exports; the computation is otherwise the same. The output is deterministic for the pinned toolchain: Manifest identifiers are fixed, protobuf maps sorted, and the conversion date coremltools records in each model.mlmodel's metadata (com.github.apple.coremltools.conversion_date) is pinned to 2026-11-01, the day these files were converted, so a conversion on another day writes the same bytes. Four conversions on 2026-11-01 gave byte-identical files.
• Toolchain: Python 3.12, torch 2.15.0, coremltools 9.0, transformers 5.22.0, safetensors 0.8.0, numpy 2.5.4, on macOS 27 (Apple silicon).

Notice of changes: the network weights were converted from PyTorch safetensors to Core ML and stored as fp16, and the model was split into three packages (image encoder, text encoder, detector). The image model only; SAM 3's video tracker is not included. The weights were not fine-tuned.

Files

• sam3/SAM3ImageEncoder.mlpackage: image 1×3×1008×1008 RGB → fpn0 1×256×288×288, fpn1 1×256×144×144, fpn2 1×256×72×72.
• sam3/SAM3TextEncoder.mlpackage: input_ids and attention_mask 1×32 (int32, CLIP BPE) → text 1×32×256.
• sam3/SAM3Detector.mlpackage: fpn0, fpn1, fpn2, text, text_mask → masks 1×200×288×288 (logits), boxes 1×200×4 (x0 y0 x1 y1 in 0…1), logits 1×200, presence 1×1, semantic 1×1×288×288. A query's score is sigmoid(logit) × sigmoid(presence).
• sam3/vocab.json, sam3/merges.txt: the CLIP tokenizer, from upstream.
• sam3/LICENSE: the SAM License. sam3/README.md: the card the conversion writes (Solume checks its licence line).

File	Bytes	SHA-256
sam3/README.md	915	f6ade2e46bd0435de97c634374b30d1de2fed6d004dc123d9b54723cd1e9e2a7
sam3/LICENSE	7,352	bec48f70bd37bf8280a9d1ebf01642d26086f6122ba735baa08fe03c5a6e7448
sam3/merges.txt	524,619	9fd691f7c8039210e0fced15865466c65820d09b63988b0174bfe25de299051a
sam3/vocab.json	862,328	5047b556ce86ccaf6aa22b3ffccfc52d391ea4accdab9c2f2407da5b742d4363
sam3/SAM3ImageEncoder.mlpackage/Manifest.json	617	a2576a37183e5a04265456092e18271e165564bd2e6578f6091e404236f3069b
sam3/SAM3ImageEncoder.mlpackage/Data/com.apple.CoreML/model.mlmodel	591,074	2c9d8a2dbc14735d67c36396642e785bf37fe27fecec70e96de9ea1faa739377
sam3/SAM3ImageEncoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin	917,318,848	fd22df3363dbf38c44be38684de64e6e6beb970f34825cdcfedf578a581be6fa
sam3/SAM3TextEncoder.mlpackage/Manifest.json	617	a2576a37183e5a04265456092e18271e165564bd2e6578f6091e404236f3069b
sam3/SAM3TextEncoder.mlpackage/Data/com.apple.CoreML/model.mlmodel	290,587	9e2c1c825f45507fb6db0c448cbc151c64eb64492db66a0f35b86708ee645034
sam3/SAM3TextEncoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin	706,427,904	72d9d6f6e977d5289583ae3c51af2d612070ef69e349ffd015a48a8cb2cdfc52
sam3/SAM3Detector.mlpackage/Manifest.json	617	a2576a37183e5a04265456092e18271e165564bd2e6578f6091e404236f3069b
sam3/SAM3Detector.mlpackage/Data/com.apple.CoreML/model.mlmodel	519,924	0abea1f103176211826cfecbb092a416dbd771ed2206cfffda7382ecce760a02
sam3/SAM3Detector.mlpackage/Data/com.apple.CoreML/weights/weight.bin	66,783,616	b419a31bcce38f7749d8aef3cb27e7d73d83866ad6700b268d0e45bde6fc0950

Licence

Meta's SAM License; full text in LICENSE, unchanged from facebook/sam3 at the revision above. Use of these files is subject to it. This conversion is not made, affiliated with or endorsed by Meta.