An assembled checkpoint that couples the Qwen3-8B temporal backbone (Korean-capable text LLM) with the PersonaPlex audio stack (Mimi codec, depth decoder with dep_q=16, and 16 audio input embeddings), toward a Korean full-duplex spoken dialogue model.

Status: before coupling training. The backbone was swapped from Helium to Qwen3-8B, so the audio stack is not yet aligned with it — audio output is expected to be noise. The text stream is already fluent (evidence that the backbone transplant is numerically correct). Background and assembly code: https://github.com/un1876/k-moshi

Usage

# pip install "git+https://github.com/un1876/k-moshi.git@main"  (not in upstream transformers)
from transformers.models.kmoshi.modeling_kmoshi import KmoshiForConditionalGeneration
model = KmoshiForConditionalGeneration.from_pretrained("spidyun/kmoshi")

Attribution / Licenses

Backbone weights: Qwen/Qwen3-8B — Apache-2.0
Audio stack (Mimi codec, depth decoder, audio input embeddings): nvidia/personaplex-7b-v1 — NVIDIA Open Model License (+ CC-BY-4.0). See LICENSE / Notice in this repository.
The Mimi codec is byte-identical to stock Moshi (kyutai, CC-BY-4.0).
depth_decoder.text_embed_tokens is re-initialized for the new tokenizer vocab (151936) and belongs to the coupling-training stage.