EgorKodin/Qwen3.8-27B-ABLITERATED-2bit-MLX-TextOnly

Qwen3.8-27B-ABLITERATED — 2-bit MLX TextOnly

This is a 2-bit MLX TextOnly quantization of Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16, optimized for inference on Apple Silicon.

Quantization

• Format: MLX
• Quantization: 2-bit affine
• Group size: 64
• Vision tower: removed
• Runtime: mlx-lm
• Base model: Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16

The language-model weights were converted to MLX and quantized to 2-bit. The vision components of the original multimodal model are not included, making this release TextOnly and reducing memory requirements.

Usage

Install MLX LM:

pip install -U mlx-lm

Run:

mlx_lm.generate \
  --model EgorKodin/Qwen3.8-27B-ABLITERATED-2bit-MLX-TextOnly \
  --prompt "Hello! Explain quantization briefly." \
  --max-tokens 512

To disable thinking when supported by the model's chat template:

mlx_lm.generate \
  --model EgorKodin/Qwen3.8-27B-ABLITERATED-2bit-MLX-TextOnly \
  --prompt "Hello!" \
  --max-tokens 512 \
  --chat-template-config '{"enable_thinking": false}'

Notes

2-bit quantization is aggressive and prioritizes low memory usage and compact model size. Depending on the task, quality may be lower than the 3-bit or higher-bit versions.

Credits

Original model:

Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16

This repository provides an MLX quantization/conversion of the original model. Credit for the underlying model and abliteration work belongs to the original authors.