Document SnapTop model distribution and provenance

README.md

@@ -2,4 +2,46 @@
 license: other
 license_name: sam
 license_link: LICENSE
+base_model: facebook/sam3
+library_name: transformers
+pipeline_tag: mask-generation
+tags:
+  - sam3
+  - snaptop
+  - onnx
 ---
+
+# SnapTop model distribution
+
+This is a dedicated model repository for SnapTop. It is separate from the
+application source repository and other repositories owned by this account.
+
+SnapTop downloads these files for local inference. Screenshots, text prompts
+and videos stay on the user's computer. Public downloads require no Hugging
+Face account or access token.
+
+## Contents and provenance
+
+- `model.safetensors` and the processor/tokenizer files: unchanged SAM3
+  Transformers weights from [facebook/sam3](https://huggingface.co/facebook/sam3),
+  revision `3c879f39826c281e95690f02c7821c4de09afae7`.
+- `sam3_vit_h.zip`: unchanged ONNX export from
+  [vietanhdev/segment-anything-3-onnx-models](https://huggingface.co/vietanhdev/segment-anything-3-onnx-models),
+  used by SnapTop's built-in image segmentation engine.
+- `LICENSE`: the complete Meta SAM License accompanying the SAM3 materials.
+- `SHA256SUMS`: checksums of the distribution files.
+
+The weights were developed by Meta, not SnapTop. No fine-tuning or model
+conversion was performed by this repository's maintainer. Redistribution
+does not change the original SAM License or its use restrictions. ONNX
+packaging and conversion attribution belongs to the upstream converter.
+
+## Downloads
+
+SnapTop release builds pin a repository revision and verify downloaded model
+files. Each backend downloads its own format on first use and reuses the local
+cache. Installing the app alone does not download both multi-gigabyte formats.
+
+The Python engine uses the Transformers model; the built-in engine uses the
+ONNX archive. Both formats are hosted here so customers do not need to request
+access to the upstream gated repository.