license: apache-2.0
base_model: hotdogs/Qwen3.8-27B-abliterated-cyber-preview
pipeline_tag: text-generation
library_name: gguf
tags:
- gguf
- qwen3_5
- abliterated
- cybersecurity
- offensive-security
- agentic
- tool-calling
- mcp
- image-text-to-text
- llama.cpp
- speculative-decoding
datasets: - hotdogs/cyber-sft-agent-qwen38
language: - en
- zh
Qwen3.8-27B-Abliterated-Cyber-Preview — GGUF (MTP)
GGUF builds of the offensive-security / agentic tool-calling merged model
hotdogs/Qwen3.8-27B-abliterated-cyber-preview
(the cyber LoRA merged into the abliterated base at scale 1.0, with the
MTP head preserved). All files are imatrix-quantized and are ready to run
in llama.cpp with native
self-speculative (MTP) decoding.
For authorized security research and red-teaming only — this model will
not refuse and may emit exploit instructions. Use only on systems you own or
are explicitly authorized to test. Apache-2.0.
Files
| file | size | notes |
|---|---|---|
Qwen3.8-27B-abliterated-cyber-preview-MTP-bf16.gguf |
54.66 GB | master (bf16), MTP in-file |
Qwen3.8-27B-abliterated-cyber-preview-MTP-Q6_K.gguf |
22.43 GB | high quality |
Qwen3.8-27B-abliterated-cyber-preview-MTP-Q4_K_M.gguf |
16.81 GB | recommended default |
Qwen3.8-27B-abliterated-cyber-preview-MTP-IQ4_NL.gguf |
16.04 GB | IQ4_NL (imatrix) |
Qwen3.8-27B-abliterated-cyber-preview-MTP-IQ3_M.gguf |
12.77 GB | smallest 3-bit (imatrix) |
Qwen3.8-27B-abliterated-cyber-imatrix.dat |
~10 MB | imatrix calibration data |
Qwen3.8-27B-calibration-BEST.txt |
— | calibration corpus |
The quantized files were produced with an imatrix (see...-imatrix.dat) computed from the bf16 master — required for the 4-bit
IQ files and improving low-bit quality.
Usage (llama.cpp)
Serve (with MTP self-speculative decoding)
./build/bin/llama-server -m Qwen3.8-27B-abliterated-cyber-preview-MTP-Q4_K_M.gguf \
--spec-type draft-mtp --spec-draft-n-max 2 -ngl 99 -c 32768
Sweep --spec-draft-n-max (1–6) to find the fastest point on your hardware.
CLI
./build/bin/llama-cli -m Qwen3.8-27B-abliterated-cyber-preview-MTP-Q4_K_M.gguf \
-ngl 99 -cnv -c 32768
The model uses the Qwen3.5 chat template; to get tool calls, give it a
penetration-test scenario as a system/user prompt:
User: Port scan the host 203.0.113.10 and identify which services are exposed.
<tool_call>
<function=nmap>
<parameter=target>
203.0.113.10
</parameter>
<parameter=ports>
-top 1000
</parameter>
</function>
</tool_call>
Key numbers
Measured on the merged model (full details in the
parent model card):
| metric | value |
|---|---|
| Tool-call emitted (6 pentest prompts) | 6/6 · 100 % |
| Correct real-tool selection | 6/6 · 100 % (nmap, ffuf, masscan, sqlmap, wpscan, smbclient) |
| General capability (7 QA/math/code prompts) | 7/7 |
| KL divergence (base ‖ merged), base-prompts | 0.041 (base preserved) |
| KL divergence (base ‖ merged), tool-prompts | 0.808 (intended re-target) |
| MTP draft acceptance rate | 0.77 (51/66) |
| Scale (merge) | 1.0 (PEFT alpha/r = 64/32 = 2.0) |
Related
- Parent model (safetensors): hotdogs/Qwen3.8-27B-abliterated-cyber-preview
- Base model: hotdogs/Qwen3.8-27B-abliterated
- Training data: hotdogs/cyber-sft-agent-qwen38
Disclaimer
Preview for authorized security research and red-teaming. This model will not
refuse and may emit instructions for exploiting systems — use only on systems
you own or are authorized to test. Apache-2.0.