license: apache-2.0 base_model: philbert440/Qwen3.8-27B-Uncensored-Cyber tags:

gguf
llamAmpere
atx
mtp
uncensored
abliterated
cyber
pentesting
security

Qwen3.8-27B-Uncensored-Cyber-ATX-MTP-GGUF

Custom ATX GGUF quantizations of Qwen3.8-27B-Uncensored-Cyber featuring an embedded Multi-Token Prediction (MTP) draft head in blk.64, specifically optimized for high-throughput offensive-security and code inference in llamAmpere and standard llama.cpp.

Model Specialization: Offensive Security & Pentesting

Domain: Penetration testing, exploit development, malware analysis, vulnerability auditing, reverse engineering, and red-teaming.
Recipe: Built on the α=1.15 Aggressive base with a residual-cyber peel (β=1.0) applied from layer 4 onwards, preserving layers 0-3 for feature extraction and structural reasoning.
Cyber-Openness: 100/100 (Zero refusal rate on offensive security queries).
Embedded MTP Head: 15 draft tensors located in blk.64 preserved at high precision (Q8_0 / q5_0) for speculative decoding draft acceptance rates above 75%.
Custom Cyber IMatrix (i1): Calibrated against real cyber-offensive code, shell scripting, exploits, and ChatML tool-calling traffic with --parse-special.

Optimized Engine: llamAmpere (SM86 / SM89 / SM90)

Engineered for llamAmpere v0.4, specifically targeting 24GB GPUs (RTX 3090 / 3090 Ti):

Hardware-Accelerated MTP Decoding (--spec-type draft-mtp): Decodes 80-100+ tokens/s on a single RTX 3090 via self-speculative draft verification.
Recurrent State Protection: SSM vectors (ssm_alpha, ssm_beta, ssm_conv1d) are locked in Q8_0 across all quants to eliminate state drift at long context.
TurboQuant KV Cache Support: Full compatibility with -ctk turbo5 -ctv turbo4 to unlock massive context without truncation.

Available Model Files

File	Size	Target Hardware	Description
`Qwen3.8-27B-Uncensored-Cyber-ATX-i1-Q8_0.gguf`	29.0 GiB	48 GB VRAM / Dual GPU	Master reference precision.
`Qwen3.8-27B-Uncensored-Cyber-ATX-i1-Q6_K.gguf`	23.6 GiB	32 GB VRAM	High-fidelity quantization with protected SSM/MTP.
`Qwen3.8-27B-Uncensored-Cyber-ATX-i1-Q5_K_M.gguf`	20.2 GiB	24 GB VRAM (RTX 3090/4090)	Recommended Sweet Spot (Speed, Context & Precision).
`Qwen3.8-27B-Uncensored-Cyber-ATX-i1-IQ4_XS-M.gguf`	15.4 GiB	24 GB VRAM	JakeATX profile (fits full 262k context under 23GB VRAM).
`imatrix_cyber_atx.dat`	27.3 MB	Calibration data	Cyber + ChatML importance matrix.

Recommended Execution (llamAmpere on Windows PowerShell)

Clone and build the repository from GitHub: JakeATX/llamAmpere, then run:

cd \llamAmpere
$env:GGML_Q8_TURBO3_MMA_FUSED="1"
.\build-sm86\bin\Release\llama-server.exe `
  -m "Qwen3.8-27B-Uncensored-Cyber-ATX-i1-Q5_K_M.gguf" `
  -ngl 99 `
  -c 131072 `
  -b 4096 `
  -ub 1024 `
  -t 20 `
  -tb 8 `
  -fa on `
  -ctk turbo5 `
  -ctv turbo4 `
  --kv-unified `
  --prio 3 `
  --parallel 1 `
  --jinja `
  --fit off `
  --cache-prompt `
  --cache-ram 4096 `
  --spec-type draft-mtp `
  --spec-draft-n-max 4 `
  --spec-draft-p-min 0.05 `
  --spec-draft-vocab-map "llamAmpere\docs\mtp-vocab\atx_65536.txt" `
  --reasoning-format none `
  --temp 0.2 `
  --top-p 0.90 `
  --top-k 40 `
  --min-p 0.05 `
  --repeat-penalty 1.08 `
  --repeat-last-n 256 `
  --alias "Qwen3.8-27B-Uncensored-Cyber-Q5_K_M" `
  --host 127.0.0.1 `
  --port 8080

Upstream References Base Architecture & Cyber Abliteration: Qwen3.8-27B-Uncensored-Cyber High-Performance Inference Engine: llamAmpere

Model Specialization: Offensive Security & Pentesting

Domain: Penetration testing, exploit development, malware analysis, vulnerability auditing, reverse engineering, and red-teaming.

Recipe: Built on the α=1.15 Aggressive base with a residual-cyber peel (β=1.0) applied from layer 4 onwards, preserving layers 0-3 for feature extraction and structural reasoning.

Cyber-Openness: 100/100 (Zero refusal rate on offensive security queries).

Embedded MTP Head: 15 draft tensors located in blk.64 preserved at high precision (Q8_0 / q5_0) for speculative decoding draft acceptance rates above 75%.

Custom Cyber IMatrix (i1): Calibrated against real cyber-offensive code, shell scripting, exploits, and ChatML tool-calling traffic with --parse-special.

Optimized Engine: llamAmpere (SM86 / SM89 / SM90)

Engineered for llamAmpere v0.4, specifically targeting 24GB GPUs (RTX 3090 / 3090 Ti):

Hardware-Accelerated MTP Decoding (--spec-type draft-mtp): Decodes 80-100+ tokens/s on a single RTX 3090 via self-speculative draft verification.

Recurrent State Protection: SSM vectors (ssm_alpha, ssm_beta, ssm_conv1d) are locked in Q8_0 across all quants to eliminate state drift at long context.

TurboQuant KV Cache Support: Full compatibility with -ctk turbo5 -ctv turbo4 to unlock massive context without truncation.

Available Model Files

File

Size

Target Hardware

Description

Qwen3.8-27B-Uncensored-Cyber-ATX-i1-Q8_0.gguf

29.0 GiB

48 GB VRAM / Dual GPU

Master reference precision.

Qwen3.8-27B-Uncensored-Cyber-ATX-i1-Q6_K.gguf

23.6 GiB

32 GB VRAM

High-fidelity quantization with protected SSM/MTP.

Qwen3.8-27B-Uncensored-Cyber-ATX-i1-Q5_K_M.gguf

20.2 GiB

24 GB VRAM (RTX 3090/4090)

Recommended Sweet Spot (Speed, Context & Precision).

Qwen3.8-27B-Uncensored-Cyber-ATX-i1-IQ4_XS-M.gguf

15.4 GiB

24 GB VRAM

JakeATX profile (fits full 262k context under 23GB VRAM).

imatrix_cyber_atx.dat

27.3 MB

Calibration data

Cyber + ChatML importance matrix.

Recommended Execution (llamAmpere on Windows PowerShell)

Clone and build the repository from GitHub: JakeATX/llamAmpere, then run:

cd \llamAmpere $env:GGML_Q8_TURBO3_MMA_FUSED="1" .\build-sm86\bin\Release\llama-server.exe ` -m "Qwen3.8-27B-Uncensored-Cyber-ATX-i1-Q5_K_M.gguf" ` -ngl 99 ` -c 131072 ` -b 4096 ` -ub 1024 ` -t 20 ` -tb 8 ` -fa on ` -ctk turbo5 ` -ctv turbo4 ` --kv-unified ` --prio 3 ` --parallel 1 ` --jinja ` --fit off ` --cache-prompt ` --cache-ram 4096 ` --spec-type draft-mtp ` --spec-draft-n-max 4 ` --spec-draft-p-min 0.05 ` --spec-draft-vocab-map "llamAmpere\docs\mtp-vocab\atx_65536.txt" ` --reasoning-format none ` --temp 0.2 ` --top-p 0.90 ` --top-k 40 ` --min-p 0.05 ` --repeat-penalty 1.08 ` --repeat-last-n 256 ` --alias "Qwen3.8-27B-Uncensored-Cyber-Q5_K_M" ` --host 127.0.0.1 ` --port 8080

Upstream References Base Architecture & Cyber Abliteration: Qwen3.8-27B-Uncensored-Cyber High-Performance Inference Engine: llamAmpere