Not to mention that huggingface-cli does not exist since more than a year!

also:

0. Install hf

pip install hf

1. Download GGUF weights

hf download OpenMOSS-Team/MOSS-TTS-GGUF --local-dir weights/MOSS-TTS-GGUF

2. Download ONNX audio tokenizer (companion weights)

hf download OpenMOSS-Team/MOSS-Audio-Tokenizer-ONNX --local-dir weights/MOSS-Audio-Tokenizer-ONNX

this downloads 66GB of USELESS data since the model is not even 7 gb

Hey — I hit almost exactly the same two problems (full-repo download, and
looking for a torch-free path) while getting this running myself, so
sharing what worked in case it saves the frustration:

The 66GB/hundreds-of-GB download: hf download <repo> --local-dir ...
with no filter pulls down every file in the repo — for MOSS-TTS-GGUF
that means all five quantization levels (Q4_K_M through F16, ~42GB
combined) plus every variant's embeddings/lm_heads. You only need one
GGUF file. Either:

hf download OpenMOSS-Team/MOSS-TTS-GGUF --local-dir weights/MOSS-TTS-GGUF --include "MOSS_TTS_Q5_K_M.gguf" --include "embeddings/*" --include "lm_heads/*" --include "tokenizer/*"

or, if --include misbehaves for you (it did for us in a couple of
cases), the Python API is more reliable:

from huggingface_hub import snapshot_download
snapshot_download("OpenMOSS-Team/MOSS-TTS-GGUF", local_dir="weights/MOSS-TTS-GGUF",
                   allow_patterns=["MOSS_TTS_Q5_K_M.gguf", "embeddings/*", "lm_heads/*", "tokenizer/*"])

Same idea for MOSS-Audio-Tokenizer-ONNX — only grab encoder.onnx* /
decoder.onnx*, not every precision variant.

Also: pip install hf isn't the right package — there's an unrelated
hf on PyPI. The official way to get the hf CLI is:

pip install -U huggingface_hub

On the "lean, no-pytorch pipeline" ask — it already exists:
moss_tts_delay.llama_cpp with the llama-cpp-onnx extra
(pip install -e ".[llama-cpp-onnx]") is fully torch-free at runtime —
llama.cpp for the text backbone, ONNX Runtime for the audio tokenizer.
No PyTorch involved. The one non-obvious gotcha: moss_audio_tokenizer
is a git submodule of this repo, not a separately pip-installable
package —

git submodule update --init --recursive

from the repo root, not git clone + pip install -e . of the
standalone OpenMOSS/MOSS-Audio-Tokenizer repo (that one pulls in
torch==2.7.0 as a hard dependency and still won't expose the right
import path anyway).

We wrote up the full CPU-only build (including the Windows-specific
parts, if relevant to anyone here) in more detail in
discussion #5 —
might save a few more hours for anyone stuck on the same points.