license: apache-2.0
base_model: ressl/Qwen3.8-27B-uncensored
base_model_relation: quantized
library_name: transformers
pipeline_tag: text-generation
language: [en]
tags: [gguf, llama-cpp, uncensored, abliterated, qwen3_5, hybrid]
Qwen3.8-27B uncensored (GGUF)
TL;DR: abliterated Qwen3.8-27B as a GGUF ladder (Q8_0, Q6_K, Q5_K_M,
Q4_K_M) plus a BF16 mmproj for the vision tower. Hard refusals on the
1120-prompt cross-eval: Q8_0 0/1120, Q4_K_M 0/1120 (BF16
master 0/1120, base model 804/1120). Built for security
research, red-teaming and penetration-testing workflows; it will comply with
requests a stock model refuses. Use responsibly.
Facts & figures
| Base model | ressl/Qwen3.8-27B-uncensored (BF16 abliterated master) |
| Method | BF16 GGUF conversion (llama.cpp) plus K-quant ladder; mmproj BF16 export for the vision encoder |
| Files | Q8_0 29.0 GB, Q6_K 22.4 GB, Q5_K_M 19.5 GB, Q4_K_M 16.8 GB, mmproj BF16 0.9 GB |
| Context length | 262,144 tokens native |
| Hardware | 1x NVIDIA RTX PRO 6000 Blackwell 96 GB (SM120) |
| Toolchain | llama.cpp (master, CUDA build for SM120), transformers 5.15.0 |
Refusal behavior (measured, hard markers only)
| Eval set | Prompts | Base | BF16 master | Q8_0 | Q4_K_M |
|---|---|---|---|---|---|
| JailbreakBench | 100 | 71 | 0 | 0 | 0 |
| tulu-harmbench | 320 | 232 | 0 | 0 | 0 |
| HarmfulQA | 300 | 207 | 0 | 0 | 0 |
| LLM-LAT | 300 | 225 | 0 | 0 | 0 |
| mlabonne harmful | 100 | 69 | 0 | 0 | 0 |
| TOTAL | 1120 | 804 | 0 | 0 | 0 |
Run it with llama.cpp (validated)
llama-server -m Qwen3.8-27B-uncensored-Q4_K_M.gguf \
--mmproj mmproj-Qwen3.8-27B-uncensored-BF16.gguf \
--jinja -ngl 99 -c 131072 --port 30003
Quality & limitations
- Quantization amplifies residual refusals: Q8_0 is closest to the BF16
master, Q4_K_M drifts a few points higher (see table). - The BF16 GGUF (54.6 GB) is not uploaded because Hugging Face rejects single
files above 50 GB; use Q8_0 for near-BF16 behaviour. - MTP speculative decoding is not wired into this export.
Provenance & reproducibility
BF16 master: abliterated Qwen/Qwen3.8-27B
with norm-preserving biprojection over 128 residual-writing matrices
(pipeline in the master repository). Conversion: convert_hf_to_gguf.py
(BF16 plus mmproj), quantization: llama-quantize. Validation: served with
llama.cpp and re-evaluated on the house cross-eval before upload.
License & credits
Apache-2.0 (inherited from Qwen3.8-27B).
Abliteration, conversion, quantization and validation by Robert Ressl
(Hugging Face -
Website -
LinkedIn).
Note: the BF16 GGUF (54.6 GB) exists locally but is not uploaded because Hugging Face rejects single files above 50 GB; Q8_0 is the highest-fidelity uploaded build. All evaluation reports in this repository were produced on the same protocol as the BF16 master (temperature 0, enable_thinking=False, hard decline markers in the first 25 words).