Qwen3.8-27B-Abliterated-MLX-6bit

Qwen3.8-27B Abliterated MLX 6-bit

An unofficial MLX 6-bit derivative of Qwen/Qwen3.8-27B. The original model is by Qwen; the MLX conversion, refusal-direction experiment, and validation were performed by PocketAI Model Lab.

This repository deliberately keeps the upstream model name first. PocketAiHub identifies the publisher of this derivative, not the creator of Qwen3.8.

Important safety notice

This checkpoint has been modified to suppress learned refusal behavior. It may produce harmful, illegal, offensive, deceptive, or dangerously incorrect content more readily than the upstream instruction model. Abliteration is not truthfulness training, capability improvement, or a safety guarantee. Use it only where you can independently evaluate and constrain its outputs.

Format

• MLX affine 6-bit, group size 64
• 498 quantized language modules
• Vision tower retained in BF16
• Stored artifact size: 22,804,849,041 bytes (21.24 GiB)
• Validated with mlx==0.32.0 and mlx-vlm==0.6.8
• Pinned base revision: 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0

Abliteration recipe

The reusable LFM2.5 workflow was adapted to Qwen3.8's 64-layer hybrid architecture. A projected harmful-minus-harmless direction was measured from 256 length-matched prompts per class at the assistant-generation boundary.

• Direction source layer: 53
• Destination layers: 24–63
• Scale: 1.0
• Per-input-column norm preservation: enabled
• Modified residual-output matrices: 80
  • 30 gated-delta/linear-attention out_proj matrices
  • 10 full-attention o_proj matrices
  • 40 MLP down_proj matrices

The edit was applied to a separate BF16 master checkpoint, and this artifact was derived from that master. The official source checkpoint was not modified in place. See abliteration-manifest.json.

Validation

Across the 200 behavioral generations, 200 reached the 128-token ceiling and 0 completed naturally. The aggregate contained 0 explicit refusals. This therefore measures early explicit refusal behavior, not whether long answers naturally reach EOS. The refusal scorer is phrase-based and cannot establish universal compliance or answer quality. Machine-readable results are in validation-summary.json.

4K performance

On an Apple M5 Max with 128 GB unified memory, batch size 1, temperature 0, seed 0, and thinking disabled:

• 4,105 prompt tokens
• 548.2 prompt tok/s
• 24.7 generation tok/s
• 7.87 seconds end to end
• 29.54 GB peak MLX memory

This is one local run, not a cross-machine performance guarantee.

Load with MLX-VLM

python -m pip install "mlx==0.32.0" "mlx-vlm==0.6.8"

from mlx_vlm import generate, load
from mlx_vlm.prompt_utils import apply_chat_template

repo_id = "PocketAiHub/Qwen3.8-27B-Abliterated-MLX-6bit"
model, processor = load(repo_id)

prompt = apply_chat_template(
    processor,
    model.config,
    "Explain why seasons occur.",
    num_images=0,
    enable_thinking=False,
)
result = generate(
    model,
    processor,
    prompt,
    max_tokens=256,
    temperature=0.0,
    enable_thinking=False,
)
print(result.text)

Image and video inputs use the normal mlx_vlm.generate media arguments.

License and attribution

The base model is Apache 2.0 licensed. See LICENSE and the official Qwen model card.