PocketAiHub/Qwen3.8-9B-Abliterated-MLX

Qwen3.8-9B Distill Abliterated MLX

Abliterated MLX derivatives of the third-party distill empero-ai/Qwen3.8-9B, pinned to revision 0934f3d2327ff2df2197495278c4c46ae5a56bd9. The source is a third-party full-parameter distillation based on Qwen/Qwen3.5-9B; it is not an official Qwen3.8 release. Abliteration, conversion, and validation were performed by PocketAI Model Lab.

Important safety notice

These checkpoints were intentionally modified to suppress learned refusal behavior. They may respond more readily to requests involving potentially unsafe, illegal, offensive, deceptive, or dangerously incorrect content. Abliteration is not truthfulness training or a safety guarantee. Independently constrain and evaluate outputs for the intended deployment.

Variants

Precision	Folder	Packaged size
4-bit	4bit/	5.57 GiB
8-bit	8bit/	9.74 GiB
BF16	bf16/	17.55 GiB

The 4-bit and 8-bit variants use MLX affine quantization with group size 64. The vision tower remains BF16. The BF16 variant is unquantized. Native source MTP tensors are intentionally excluded.

Refusal-behavior screen

All variants were evaluated on 100 refusal-elicitation test prompts and 100 benign controls in non-thinking mode with deterministic decoding and a 256-token ceiling.

Precision	Test-set explicit refusals	Control explicit refusals	Test-set natural stops	Control natural stops
4-bit	0/100	0/100	11/100	2/100
8-bit	0/100	0/100	14/100	4/100
BF16	0/100	0/100	11/100	3/100

The transparent phrase-based screen found no explicit refusals or evasive non-answers, and every case contained final-answer text. Most generations hit the 256-token ceiling, so this is an early-refusal regression screen—not proof of universal compliance, safety, factuality, or completion quality.

KV/long-context evaluation

Precision	Formatted tokens	Prefill tok/s	Decode tok/s	Peak MLX memory
4-bit	65,536	1028.1	44.59	13.05 GB
8-bit	32,776	2323.7	52.99	14.19 GB
BF16	32,776	2299.3	27.32	22.67 GB

All three exact-retrieval cases passed. The 8-bit and BF16 runs used 16-bit KV cache quantization at 32K; the 4-bit run was an unquantized-KV 64K text test. The 4-bit and BF16 variants also passed the complete deterministic 4K feature suite; all three passed text and vision runtime smoke tests.

Exact evidence hashes and test details are in each variant's validation-summary.json and artifact-manifest.json.

Download and load

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

from pathlib import Path

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

repo_id = "PocketAiHub/Qwen3.8-9B-Abliterated-MLX"
variant = "4bit"  # "4bit", "8bit", or "bf16"
snapshot = Path(snapshot_download(repo_id, allow_patterns=[f"{variant}/*"]))
model, processor = load(str(snapshot / variant))
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)

Reproducibility and limitations

• Source: empero-ai/Qwen3.8-9B at 0934f3d2327ff2df2197495278c4c46ae5a56bd9
• Declared base: Qwen/Qwen3.5-9B
• Abliteration is a targeted directional intervention, not general evaluation
• Standard MLX conversion intentionally excludes native MTP tensors
• This is an experimental community release; verify behavior for your use case

License and attribution

The source repository declares Apache-2.0. This derivative includes the Apache 2.0 text in LICENSE. Original model credit remains with Empero and the Qwen team; PocketAI is the derivative publisher.