- model/: bf16 model.safetensors + tokenizer/config files VERBATIM (sha256 vs upstream LFS)
- codec/: MOSS-Audio-Tokenizer re-sharded by tensor into decoder.safetensors
(decoder.* + quantizer.*) and encoder.safetensors (encoder.*); every tensor
byte-identical to upstream (per-tensor sha256 in codec/sha256.json)
- HF remote-code .py files dropped: MAESTRO vendors the GitHub moss_tts_delay/
graph + the codec (ai/moss_tts_vendor), never trust_remote_code
Loaded by MAESTRO via backend/ai/models/moss_voicegenerator.py (decoder-only codec).
