Hy-Nomas MLX 0.1.0 — independent Apple Silicon research preview
Copyright 2026 the Hy-Nomas MLX contributors.

Based on Tencent Hy-Nomas. Copyright (C) 2026 Tencent. All rights reserved.
Upstream: https://github.com/Tencent-Hunyuan/Hy-Nomas
Reference commit: 8f6c67db6eca400f53050c58e22f0ef1b5d22a9b
Full original license/third-party notices: upstream/LICENSE

Modifications and additions in this distribution:
- hynomas_mlx/: Apple MLX/Metal inference implementation of the released
  transformer, PixelUMM decoder and text conditioning; bounded native tensor
  loading; fused source-weight reconstruction and BF16 operations; sequential
  process execution for memory release. This is not upstream CUDA inference.
- experiments/camera_demo/: CPU camera preparation and offline planned-input
  harness adapted from the released camera equations. No live input service.
  The release copy adds an ffmpeg/ffprobe availability check before generation;
  --plan-only remains available without media tools. Inference math is unchanged.
- scripts/, tests/, prompts/, setup.sh and run.sh: independent inference,
  validation, packaging and example helpers. setup.sh/download_models.py in
  this release add platform/weight checks and narrow the download allowlist.
- upstream/: selected unmodified reference files, retained for parity tests.

PixelUMM source is derived from nv-tlabs/PixelUMM (Apache-2.0 code).
Released PixelUMM checkpoint weights and derivatives have the NVIDIA One-Way
Noncommercial research/evaluation notice. See LICENSING.md and included license.

Text conditioning uses MLX-community's Qwen2.5-VL-7B-Instruct 8-bit conversion
of Qwen's original model. See the retained model card and applicable license.

No weights, third-party dependency binaries, private configuration, tunnels,
credentials, caches, generated videos or prior run data are distributed here.
