botp/supergemma4-26b-abliterated-multimodal-mlx-4bit

Support ongoing open-source work: ko-fi.com/jiunsong

SuperGemma4-26B-Abliterated-Multimodal MLX 4bit

This is the lighter-weight MLX build of Jiunsong/supergemma4-26b-abliterated-multimodal.

It preserves multimodal behavior while reducing local storage and memory demand for Apple Silicon setups that want a smaller package.

Important note on the Hugging Face size badge

If the Hub UI shows this repo as a smaller class such as 5B or 8B, that is a Hub-side auto-inference artifact from the exported MLX quantized config.

This repo is still a quantized release of the full SuperGemma4-26B-Abliterated-Multimodal line derived from the Gemma 4 26B-A4B multimodal family. The smaller badge does not mean the model was accidentally converted into a different 5B or 8B model.

Why this variant

• Smaller MLX footprint for local use
• Keeps text + vision support
• Preserves the abliterated / low-refusal behavior of the main release
• Good option when you want better fit on-device without dropping multimodality
• Verified with both text-only and image-grounded prompts

Validation

• Text check: returned READY
• Image check: returned red for a solid red test image
• Disk footprint: about 15 GB

Recommended use

Pick this version when you want a smaller MLX package and are willing to trade a bit of precision for a lighter local deployment.

Quick start

python3 -m mlx_vlm.server \
  --model /absolute/path/to/supergemma4-26b-abliterated-multimodal-mlx-4bit \
  --host 127.0.0.1 \
  --port 8091

from mlx_vlm import load

model, processor = load("/absolute/path/to/supergemma4-26b-abliterated-multimodal-mlx-4bit")
print("Loaded.")