supergemma4-e4b-abliterated MLX 6-bit

This repository contains an MLX-LM conversion of Jiunsong/supergemma4-e4b-abliterated.

Conversion Details

Original model: Jiunsong/supergemma4-e4b-abliterated
Model family: SuperGemma4
Source model type: gemma4
Model size: 7,518,068,992 parameters
Quantization: MLX-LM affine quantization
Bits: 6-bit
Group size: 64
Local MLX folder size at upload time: 5.71 GiB
Local safetensors weight size at upload time: 5.68 GiB

This Gemma conversion follows the MLX-LM Gemma 4 shared-KV topology and uses non-strict checkpoint loading so extra HF tensors outside that topology are discarded during conversion.

For mlx-swift compatibility, per_layer_model_projection was left unquantized while the rest of the eligible linear layers were quantized.

For oMLX / mlx-swift chat-template compatibility, the published MLX package uses a simplified Gemma 4 chat template based on the <|turn>role\n...<turn|> format. The original upstream chat_template.jinja is preserved as chat_template.original.jinja.

Abliteration credit: the abliterated source model was published by Jiunsong. This repository only provides the MLX conversion and quantization.

Usage

mlx_lm.generate --model DreamFoundries/supergemma4-e4b-abliterated-6bit --prompt "Hello" --max-tokens 64

Benchmarks

No comparative benchmarks have been run yet. The repository does not currently provide quality, speed, memory, or benchmark comparisons against the original weights or other quantizations.

License

This is a converted/quantized derivative of the original model. Please refer to the original model repository for the upstream license and usage terms: https://huggingface.co/Jiunsong/supergemma4-e4b-abliterated