MOSS-SoundEffect-v2-CoreAI

MOSS SoundEffect v2 for Core AI

Experimental Core AI resource bundle and native Swift runtime for OpenMOSS-Team/MOSS-SoundEffect-v2.0.

The downloadable .sfxasset contains three materialized .aimodel components and the tokenizer required to generate 48 kHz mono sound effects locally on an Apple silicon Mac. Source, conversion recipes, tests, and the Swift runtime are available at caguilar-dev/moss-soundeffect-coreai.

Status

This is an experimental macOS 27 / Xcode 27 release, not a stable model release. The bundle preserves the preferred behavior of the published MLX model by reconstructing its exact effective DiT weights in the Core AI-authored graph. Targeted café and rain listening was accepted. The complete prompt suite and blinded review remain open.

Instruments measured the current denoiser on the GPU. This release does not claim Neural Engine execution.

Bundle

MOSS-SoundEffect-v2.sfxasset/
├── manifest.json
├── integrity.json
├── LICENSE
├── NOTICE
├── Models/
│   ├── TextEncoder.aimodel/
│   ├── Denoiser.aimodel/
│   └── DACDecoder.aimodel/
├── Tokenizer/
└── Validation/validation.json

• Text encoder: uncompressed Float16 Qwen3-1.7B encoder
• Denoiser: uncompressed Float16 Core AI graph using effective MLX affine-INT4 weights
• Decoder: uncompressed Float16 DAC decoder
• Output: mono PCM audio at 48 kHz, up to 30 seconds

Download

huggingface-cli download \
  caguilar51/MOSS-SoundEffect-v2-CoreAI \
  --local-dir MOSS-SoundEffect-v2-CoreAI

Run

Build the Swift CLI from the GitHub repository, then run:

Runtime/.build/release/moss-sfx-coreai \
  --bundle MOSS-SoundEffect-v2-CoreAI/MOSS-SoundEffect-v2.sfxasset \
  --prompt "background conversation in a coffee shop during lunch" \
  --duration-seconds 10 \
  --output coffee.wav

The .sfxasset suffix is this project's resource-bundle convention; it is a directory, not an Apple model-file format.

Validation

• Matched denoiser pass against MLX: 58.75 dB PSNR
• Swift/Core AI stride correction: 20-step latent cosine improved to 0.9988
• Targeted café/rain listening: accepted
• Materialized .sfxasset Swift API smoke test: passed
• Measured compute placement: GPU
• Full exact-candidate suite and blinded review: pending

See the GitHub repository's validation document for methodology, accepted and rejected experiments, and limitations.

License and attribution

Apache-2.0. The source model, published MLX conversion, and Qwen3-1.7B components are Apache-2.0. The source repository also contains a patch against Apple's BSD-3-Clause coreai-models runtime; Apple developer tools are not included in this model repository.