# Bundled Model Pack Notices

The model packs below are redistributed unmodified from immutable upstream
revisions. Momentum cinemaOS adds only packaging metadata, checksum validation,
local loading, and application-side pre/post-processing.

## Real-ESRGAN ×4, BSRGAN ×4 and MPRNet Denoise

Core ML conversions are redistributed unmodified from TheMurusTeam:

- `coreml-upscaler-realesrgan512`, revision `8584e3fa79317be747c5edbad519e489c989e434`
- `coreml-upscaler-bsrgan`, revision `a96f019776dba50d0f46a658d549d7ceb624fe41`
- `coreml-upscaler-MPRNetDenoising`, revision `b80fe88c3dd365200802594575fc9b539709bc43`

The upstream model cards specify CreativeML OpenRAIL-M. The complete license,
including its Attachment A use restrictions, is supplied in
`Licenses/CreativeML-OpenRAIL-M.txt`. These restrictions apply to use and
redistribution of these models. They are not unrestricted-use model weights.
Release review must ensure the application agreement incorporates the required
use restrictions before public distribution.

The archives are part of the verified application payload. Extraction and Core ML
compilation occur only in writable user storage, never inside the signed app.

## Depth Anything V2 Small P6

- Source: `apple/coreml-depth-anything-v2-small`
- Revision: `cfef6f6f2a70783dedc0bfae40cecbc2052285d3`
- Artifact: `DepthAnythingV2SmallF16P6.mlpackage`
- License: Apache License 2.0
- Original work: Depth Anything V2, Lihe Yang and contributors
- Core ML conversion: Apple

Only the Small variant is included. The Base, Large, and Giant weights are not
included because their noncommercial license is incompatible with this public
release.

## Parakeet TDT 0.6B v3 Core ML

- Source: `FluidInference/parakeet-tdt-0.6b-v3-coreml`
- Revision: `7dd20fe6b1797d35f5e3307e8b1732d9a178edfe`
- Artifacts: production v3 preprocessor, int8 encoder v2, decoder, joint v3,
  and vocabulary
- Weight license: Creative Commons Attribution 4.0 International
- Original work: NVIDIA Parakeet TDT 0.6B v3
- Core ML conversion: FluidInference

Attribution: NVIDIA and FluidInference. The files are not modified. The
application identifies the model and license in its acknowledgements.

## Sortformer High Context v2.1, 6-bit Core ML

- Source: `FluidInference/diar-streaming-sortformer-coreml`
- Revision: `ae9a27ab45dc0aa3abede7d2d6bad2b7a69aa6d1`
- Artifact: `v3/palettized/SortformerNvidiaHigh_v2.1.mlmodelc`
- Repository card license: Creative Commons Attribution 4.0 International
- Upstream model terms: NVIDIA Open Model License
- Original work: NVIDIA Sortformer
- Core ML conversion and palettization: FluidInference

Both upstream notices are shipped because the repository metadata and model
documentation name different permissive terms. The more restrictive combined
set of attribution and notice obligations is followed.

## LaMa Core ML

- Source: `mlboydaisuke/LaMa-CoreML`
- Revision: `5ed76e3799ab4cad31381750d29880c267477e18`
- Artifact: `LaMa.mlpackage`
- License: Apache License 2.0
- Original work: Resolution-robust Large Mask Inpainting with Fourier
  Convolutions, Roman Suvorov and contributors
- Core ML conversion: Daisuke Majima

LaMa is exposed only as a non-destructive derived-media job. Original media is
never overwritten.

## FluidAudio

- Source: `FluidInference/FluidAudio`
- Revision: `8a2bf2c074a227b25baef2185a00faa3f6865d10`
- License: Apache License 2.0
- Use: Apple-native Parakeet TDT v3 and Sortformer execution

Momentum cinemaOS supplies immutable, checksum-verified model directories to
FluidAudio directly. Its Hugging Face download helpers are not called by the
release runtime. The complete Apache-2.0 text is shipped as
`Licenses/FluidAudio-Apache-2.0.txt`.

## NemoTextProcessing 0.3.0

FluidAudio links the checksum-pinned `NemoTextProcessing.xcframework` 0.3.0
binary from `FluidInference/text-processing-rs`. It is Apache-2.0 and contains
an Apache-2.0 derivative of NVIDIA NeMo Text Processing plus permissively
licensed Rust dependencies. The upstream binary-target notice, including the
NVIDIA, rustfst, and flate2 attribution, ships as
`Licenses/NemoTextProcessing-NOTICE.md`; the full Apache and MIT texts ship
beside it.