Qwen3-4B-Instruct-MLX-4bit

This model is a 4-bit quantized version of rautaditya/Qwen3-4B-Instruct-2507-heretic-1 for Apple Silicon using MLX.

Features

• 4-bit Quantization: Affine mode with group size 64 for efficient inference.
• MLX Support: Native support for Mac computers with the M-series chips.
• Chat Ready: Optimized for instruction-following and thinking tasks.

Usage

Installation

Ensure you have mlx-lm installed:

pip install mlx-lm

Inference

You can run the model using the mlx-lm command line or in your Python code.

Example Python Script

from mlx_lm import load, generate

model, tokenizer = load("rautaditya/Qwen3-4B-Instruct-MLX-4bit")

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "What are some good features of MLX?"}],
    tokenize=False,
    add_generation_prompt=True,
)

response = generate(model, tokenizer, prompt=prompt, verbose=True)
print(response)

Details

• Base Model: rautaditya/Qwen3-4B-Instruct-2507-heretic-1
• Quantization: 4-bit Affine (group_size=64)
• Framework: MLX-LM