MOSS-SoundEffect — MLX selective 4-bit (Swift-compatible)

Selective 4-bit (affine, group size 64) MLX conversion of OpenMOSS-Team/MOSS-SoundEffect, a sound-effect / ambient generator on the MossTTSDelay architecture.

Selective quant: lm_heads and the input embeddings are kept dense (full precision); only the transformer body is 4-bit. This matches the module layout in mlx-audio-swift's MossTTSModel, which defines those layers as plain (non-quantized) modules — a fully-4-bit checkpoint fails to load there with a size mismatch. ~16 GB bf16 → ~5.4 GB.

Usage (Swift)

Requires the patched fork nullic/mlx-audio-swift (branch trails): adds float32 audio-token sampling + quantized-checkpoint loading to MossTTSModel. Pair with the original audio tokenizer OpenMOSS-Team/MOSS-Audio-Tokenizer.

Sampling: audio temperature 1.7, top-p 0.8, top-k 25 (the upstream reference settings; the fork's float32 sampling is what makes 1.7 coherent rather than noise).

License

Apache-2.0, inherited from the base model. Conversion only; no weights authored here.

Downloads last month: 74

Safetensors

Model size

8B params

Tensor type

BF16

U32

MLX

Hardware compatibility

4-bit

Model tree for dpetrusevich/MOSS-SoundEffect-MLX-4bit

Base model

OpenMOSS-Team/MOSS-SoundEffect

Quantized

(3)

this model