ConvMixer is a simple yet effective vision architecture that combines large-kernel depthwise convolutions for spatial mixing with pointwise convolutions for channel mixing, achieving transformer-like performance with minimal complexity.

Original paper: Patches Are All You Need? ConvMixer

This model uses the ConvMixer-768/32 variant, which processes 32×32 patches with 768 feature channels, providing strong accuracy while remaining computationally efficient. It is well suited for image classification tasks where simplicity, speed, and high accuracy are desired, and can serve as a lightweight backbone for research or prototyping.

Model Configuration: