Qwen3.8-27B-Whitehat (Quadux)
Copyright 2026 Quadux IT GmbH

This product includes software and model weights developed by third parties,
redistributed here in modified form under the Apache License, Version 2.0.

------------------------------------------------------------------------------
Base model
------------------------------------------------------------------------------
Qwen3.8-27B
Copyright the Qwen team, Alibaba Cloud.
Licensed under the Apache License, Version 2.0.
Source: https://huggingface.co/Qwen/Qwen3.8-27B

The model weights distributed here are a derivative work of Qwen3.8-27B.

------------------------------------------------------------------------------
Modifications by Quadux IT GmbH (Apache-2.0 Section 4(b))
------------------------------------------------------------------------------
The following changes were made to the base model to produce this derivative
work:

  * Selective behavioural fine-tuning (multimodal LoRA, merged into the
    weights) that opens computer-/network-security and offensive-security
    tasks (exploit development, malware and C2 analysis, reverse engineering,
    phishing-infrastructure and awareness testing, copyright-/DRM- and other
    legal-restriction bypass for feasibility and security research) while
    RETAINING refusal of physically harmful content (weapons, explosives,
    drug synthesis, CBRN, violence) and CSAM. The boundary is enforced
    language-independently and across the text and vision (image) input paths.

  * Multi-Prediction-Token (MTP) head preserved and re-grafted after the
    transformers save step (block blk.64).

  * GGUF conversion and quantization (imatrix-assisted K-quant ladder plus an
    Unsloth-Dynamic UD-Q4_K_XL per-tensor recipe). The vision projector
    (mmproj-F16.gguf) is the unmodified original.

The vision tower ("visual") was frozen during fine-tuning and is unchanged
from the base model.

------------------------------------------------------------------------------
Third-party quantization scheme (attribution)
------------------------------------------------------------------------------
The UD-*_K_XL GGUFs apply the Unsloth "Dynamic 3.0" per-tensor quantization
type maps. The tensor-by-tensor type assignments were derived from Unsloth's
community GGUF builds of the Qwen3.8-27B base model (UD-Q2_K_XL through
UD-Q8_K_XL variants) and applied - via llama.cpp's --tensor-type-file - to our
own fine-tuned weights, together with an importance matrix we computed
ourselves. No Unsloth weights are redistributed here; only the quantization
recipe (the type maps) was reused.

Unsloth AI - https://huggingface.co/unsloth - https://unsloth.ai
Distributed under the Apache License, Version 2.0.
With thanks to the Unsloth team for the Dynamic-quantization methodology.

------------------------------------------------------------------------------
Calibration data
------------------------------------------------------------------------------
The importance matrix (imatrix) used for the lower quantization levels was
computed with the publicly available "calibration_datav3" text corpus
(bartowski). No end-user or customer data was used.

------------------------------------------------------------------------------
This NOTICE file is provided for informational purposes only and does not
modify the License. See the LICENSE file for the full Apache License 2.0 terms.

Use of this model is subject to the Disclaimer and Terms of Use — see DISCLAIMER.md and the model card. / Die Nutzung unterliegt dem Haftungsausschluss — siehe DISCLAIMER.md und die Model-Card.