hans00/MOSS-TTS-Nano-GGUF

MOSS-TTS-Nano GGUF (codec only)

GGUF conversion of the MOSS-Audio-Tokenizer-Nano codec used by `OpenMOSS-Team/MOSS-TTS-Nano`, runnable via codec.cpp.

⚠️ The LLM-part is not converted yet — see Status.

Files

Codec-part (MOSS-Audio-Tokenizer-Nano, moss_audio arch, 16 RVQ codebooks × 1024, 48 kHz stereo)

codec[-<quant>].gguf

Status

The MOSS-TTS-Nano LLM-part is a custom `MossTTSNanoForCausalLM` architecture that stock llama.cpp doesn't load:

• —GPT-2 backbone with RoPE (position_embedding_type: "rope", rope_base: 10000) — llama.cpp's gpt2 arch only handles learned absolute positions, not RoPE
• —16 RVQ codebooks (1024 entries each) emitted per audio frame
• —Global + local transformer: the GPT-2 global produces a hidden state per timestep and a 1-layer local transformer expands it into the 16 codebook predictions

A working LLM-part GGUF would need a custom moss_tts_nano architecture inside llama.cpp (model loader, graph builder for global + local transformer, multi-codebook output handling). That's substantial work — comparable to or larger than implementing Chatterbox-T3.

For now this repo only ships the codec; the LLM-part will follow if/when the architecture lands in llama.cpp.

Notes

• —Source weights: `OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano`
• —Full upstream Python pipeline: `OpenMOSS-Team/MOSS-TTS-Nano`