Distillation — generating high-quality training data for the smaller student model
Iterative retraining — serving as the starting point for the next training cycle
High-quality inference — when maximum accuracy is needed (at the cost of speed/memory)

For deployment and everyday use, prefer the distilled 8B student model: ARO-Lang/aro-coder-6bit


Architecture	Qwen3 30B MoE (3.3B active parameters)
Base model	mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit
Quantization	4-bit (MLX)
Size	~16 GB
Training source	dpo

Usage

Distilled student: <a href="https://huggingface.co/ARO-Lang/aro-coder-6bit" target="_blank" rel="noopener noreferrer" class="text-blue-400 hover:text-blue-300 underline underline-offset-2 transition-colors">ARO-Lang/aro-coder-6bit
Website: <a href="https://arolang.github.io/aro/" target="_blank" rel="noopener noreferrer" class="text-blue-400 hover:text-blue-300 underline underline-offset-2 transition-colors">arolang.github.io/aro
GitHub: <a href="https://github.com/arolang/aro" target="_blank" rel="noopener noreferrer" class="text-blue-400 hover:text-blue-300 underline underline-offset-2 transition-colors">github.com/arolang/aro
Language Guide: <a href="https://github.com/arolang/aro/wiki" target="_blank" rel="noopener noreferrer" class="text-blue-400 hover:text-blue-300 underline underline-offset-2 transition-colors">Wiki

python

from mlx_lm import load, generate
model, tokenizer = load("ARO-Lang/aro-teacher-30b-4bit")

Or as a base for continued fine-tuning:

python -m mlx_lm lora --model ARO-Lang/aro-teacher-30b-4bit --data ./train_data --train

License

MIT License