strangecreator/Qwen3.8-27B-OS-Software-Heretic-v3-MLX-8bit-g64-affine

Qwen3.8 27B · Abliterated by OS-Software · 8-bit MLX

Converted from OS-Software/Qwen3.8-27B-Uncensored-Heretic-v3 for MLX on Apple Silicon.

Editions: 4-bit · 8-bit · BF16 source · Conversion record

General properties

Property	Value
Weights	29.50 GB / 27.48 GiB · 6 safetensors shards
Quantization	8-bit affine · group size 64
Scales, biases and unquantized weights	BF16
Vision encoder	Included in BF16
Configured context	262,144 tokens, inherited from the source
Source revision	e5fe37a1

Download for oMLX

hf download strangecreator/Qwen3.8-27B-OS-Software-Heretic-v3-MLX-8bit-g64-affine \
  --local-dir "$HOME/.omlx/models/strangecreator/Qwen3.8-27B-OS-Software-Heretic-v3-MLX-8bit-g64-affine"

Use ~/.omlx/models as your oMLX model directory, or replace it with your configured path.

Run with mlx-vlm

pip install 'mlx==0.32.3' 'mlx-vlm==0.7.6' \
  'mlx-lm==0.32.0' 'transformers==5.17.0' 'numpy==2.3.5'

mlx_vlm.generate --model strangecreator/Qwen3.8-27B-OS-Software-Heretic-v3-MLX-8bit-g64-affine \
  --prompt 'What is 2 + 2?' --temperature 0 --max-tokens 32

For image input, add --image /path/to/image.png and use an image-related prompt.

Compatibility

The source repository contains no MTP weight tensors. The source tokenizer and chat template are included. qwen3_5 in the configuration is the architecture identifier used by this Qwen3.8 model.

Reproduce the conversion

hf download OS-Software/Qwen3.8-27B-Uncensored-Heretic-v3 \
  --revision e5fe37a1fcbd74fc9058170c6c60fe248828e205 \
  --local-dir ./source-bf16

Before conversion, extract the source's nested processor settings:

import json
from pathlib import Path

root = Path("./source-bf16")
config = json.loads((root / "processor_config.json").read_text())
for key, filename in (
    ("image_processor", "preprocessor_config.json"),
    ("video_processor", "video_preprocessor_config.json"),
):
    (root / filename).write_text(json.dumps(config[key], indent=2))

mlx_vlm.convert --hf-path ./source-bf16 --mlx-path ./mlx-model \
  --quantize --q-bits 8 --q-group-size 64 \
  --q-mode affine --dtype bfloat16

The converter uses native MLX affine quantization. Both editions were converted directly from BF16.

Tool	Version
MLX	0.32.3
mlx-vlm	0.7.6
mlx-lm	0.32.0
Transformers	5.17.0

Full provenance is in conversion.json. Transfer checksums are in SHA256SUMS.

Credits and license

Original model: Qwen/Qwen3.8-27B. Abliterated BF16 checkpoint: OS-Software/Qwen3.8-27B-Uncensored-Heretic-v3.

Apache-2.0. The source declares Apache-2.0. The included license text comes from the official Qwen base model. See LICENSE and the upstream model card for intended use and evaluation details.