groxaxo/MOSS-TTS-v1.5-Argentina-MLX-oQ5-HQ8Speech-G64

MOSS-TTS v1.5 Argentina — MLX mixed-bit

MLX deployment build of a locally trained Argentine Spanish LoRA for OpenMOSS-Team/MOSS-TTS-v1.5. The selected adapter continued the selected female checkpoint for three epochs over the male Argentine split.

Quantization uses the recipe recovered from the owner's Mac:

• —normal Qwen3 backbone matrices: Q5 affine, group size 64
• —v_proj, down_proj, and transformer blocks 0–3 and 32–35: Q6 affine, group size 64
• —embeddings, LM/audio heads, and other speech-path modules: Q8 affine, group size 64
• —norms, biases, and the external audio tokenizer remain full precision

The LoRA was merged into luisarn/MOSS-TTS-v1.5-MLX-8bit before quantization. Adapter source label: MOSS-TTS-v1.5-Argentina-combined-female-epoch1-then-male-epoch2.

Load with a compatible mlx-audio checkout containing support for MOSS-TTS v1.5 and per-layer mixed quantization entries.