selcukkubur/Confucius4-R2T2-mlx-4bit

Confucius4-R2T2, MLX 4-bit

netease-youdao/Confucius4-R2T2 converted to MLX and quantised to 4 bits (group size 64, affine), for Apple Silicon. 4.09 GB of bfloat16 becomes 1.5 GB.

Converted with mlx-audio 0.3.1:

python -m mlx_audio.convert --hf-path netease-youdao/Confucius4-R2T2 \
    --mlx-path Confucius4-R2T2-mlx-4bit \
    -q --q-group-size 64 --q-bits 4 --model-domain stt

The architecture is unchanged — Qwen3ASRForConditionalGeneration, the same one Qwen/Qwen3-ASR-1.7B uses — so it loads through existing Qwen3-ASR paths in both mlx-audio and mlx-audio-swift with no custom code.

Use

from mlx_audio.stt.utils import load_model, load_audio
model = load_model("selcukkubur/Confucius4-R2T2-mlx-4bit")
print(model.generate(load_audio("audio.wav")).text)

Notice

Any modifications made to the original model in this Derivative Work are not endorsed, warranted, or guaranteed by the original right-holder of the original model, and the original right-holder disclaims all liability related to this Derivative Work.

This is a quantisation of NetEase Youdao's Confucius4-R2T2 and is governed by the original MODEL_LICENSE. That licence is royalty-free for most users but requires a separate licence from NetEase Youdao above 100 million monthly active users or RMB 1 billion in annual revenue, and prohibits use in the high-risk scenarios it lists. Read it before use.