This repository contains a native MLX conversion of the original model using bfloat16 (BF16) precision, providing maximum inference quality while leveraging Apple's unified memory architecture.

Base Model

Base repository: Irfanuruchi/Qwen3-4B-Computer-Science
Architecture: Qwen3-4B
Format: MLX
Precision: BF16 (bfloat16)

Features

Native Apple MLX format
Optimized for Apple Silicon (M-series)
Full BF16 precision
High-quality local inference
Compatible with mlx-lm

Installation

bash

python3 -m venv .venv
source .venv/bin/activate

pip install mlx mlx-lm

Usage

bash

mlx_lm.generate \
  --model Irfanuruchi/Qwen3-4B-Computer-Science-MLX-BF16 \
  --prompt "Write a Python function that validates an IPv4 address." \
  --max-tokens 256

Model Information

Property	Value
Base Model	Qwen3-4B-Computer-Science
Precision	BF16
Format	MLX
Target Platform	Apple Silicon
Framework	MLX

License

This repository is released under the Apache 2.0 License.

The original Qwen3 model is licensed under Apache 2.0. This repository contains an MLX BF16 conversion of the original weights.

Acknowledgements

Alibaba Qwen Team
Apple MLX
Hugging Face