# Licenses and model provenance

This package is an independent Apple Silicon **research/evaluation preview**.
It distributes source code and small reference/license files, not model weights.
The setup tools download checkpoints from their original repositories.

## Source code

The MLX implementation and release helpers are provided under Apache-2.0; see
`LICENSE` and `NOTICE`. Tencent's complete license and third-party attribution
file is retained at `upstream/LICENSE`. Selected original upstream reference
files are included unchanged to support the numerical/CPU parity tests.

The MLX implementation is a port, not the original CUDA code. Its source files
describe their changed backend and behavior; `NOTICE` records the modifications
and exact upstream commit. The verified core is retained byte for byte in this
release; packaging did not change inference math or the source core checksum.

## Weights and third-party terms

Tencent's pinned [Hy-Nomas LICENSE](https://huggingface.co/tencent/Hy-Nomas/blob/bb6a8ca451bfe84947f0ea1c6d51957f983ffffe/LICENSE)
states Apache-2.0 for its released code/weights with exceptions for third-party
components. Its pinned [PixelUMM decoder source](https://github.com/Tencent-Hunyuan/Hy-Nomas/blob/8f6c67db6eca400f53050c58e22f0ef1b5d22a9b/hymm/models/vae/pixelumm_decoder/model.py)
separately identifies released PixelUMM weights and derivatives as NVIDIA
One-Way Noncommercial, for research/evaluation only.

The [official NVIDIA PixelUMM model card](https://huggingface.co/nvidia/PixelUMM)
also identifies its checkpoint as a separate artifact under the
[NVIDIA One-Way Noncommercial License](https://developer.download.nvidia.com/licenses/NVIDIA-OneWay-Noncommercial-License-22Mar2022.pdf).
A copy is included in `licenses/NVIDIA-OneWay-Noncommercial-License-22Mar2022.pdf`.
We retain these notices and do not represent the combined pipeline as having
unrestricted commercial rights. Consult the applicable licensors if commercial
use of the decoder or its derivatives is required.

The [MLX-community Qwen checkpoint](https://huggingface.co/mlx-community/Qwen2.5-VL-7B-Instruct-8bit/tree/cb8e0fbe4aa086c19a3d8aaa43e79919594611f9)
and its [Qwen2.5-VL-7B-Instruct base model](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct)
identify Apache-2.0. This package uses its 8-bit language layers for conditioning.
The pinned MLX model card is retained in `licenses/qwen-mlx-model-card.md`.

Python packages installed by pip retain their respective original licenses; they
are not bundled in this source ZIP. Their distribution metadata contains their
license records. Model downloading and use remain subject to the source terms.

No endorsement by Tencent, NVIDIA, Qwen, MLX-community, or Apple is implied.