goodman117/moshi-ita-A2

i-moshi — Italian Moshi (LoRA)

This model is useless because it has an english brain but got the italian accent from the finetuning. It's my 5th try and for the first time is printing out words inside the moshi GUI. All the previous finetuning did not produce any word outputs. It kinda works, there is a lot of work to do. I guess you can try it for fun

LoRA finetune of kyutai/moshiko for Italian full-duplex spoken dialogue, in the style of J-Moshi.

Model details

• Base model: kyutai/moshiko (7B, Mimi codec @ 24 kHz)
• Method: LoRA (rank 128, scaling 2.0) via moshi-finetune
• Data: ~1,400 h of two-speaker Italian dialogue, converted to Moshi's stereo format (main speaker left, interlocutor right) with word-level aligned text, via pyannote diarization + faster-whisper.
• Training: 2 epochs (6,300 steps, batch 16, 100 s windows) on a single H100.

Limitations

Early research checkpoint: the model acquired Italian phonetics and prosody, but sentence-level coherence and turn-taking are still weak. Treat it as a research artifact, not a usable assistant.

How to run it

Inference runs with the official moshi package by Kyutai (this is a different repo from moshi-finetune, which is only needed for training). A CUDA GPU with ~16 GB of VRAM is recommended. On Windows, run everything inside WSL.

# 1. Clone the inference repo and install the moshi package
git clone https://github.com/kyutai-labs/moshi.git
cd moshi
pip install -e moshi   # the Python package lives in the moshi/ subdirectory

# 2. Download the LoRA adapter from this repo
pip install huggingface_hub
huggingface-cli download <TUO_USERNAME>/i-moshi --local-dir ./i-moshi

# 3. Launch the web UI — the LoRA adapter is merged on top of the base
#    moshiko weights at load time (base weights are auto-downloaded from HF)
CUDA_VISIBLE_DEVICES=0 python -m moshi.server \
  --lora-weight="./i-moshi/lora.safetensors" \
  --config-path="./i-moshi/config.json"

Then open http://localhost:8998 in your browser and start talking.