Qwen3-14B-Abliterated-GGUF

library_name: gguf license: apache-2.0 license_link: https://huggingface.co/Qwen/Qwen3-14B/blob/main/LICENSE pipeline_tag: text-generation base_model:

RootMonsteR/Qwen3-14B-Abliterated base_model_relation: quantized tags:
gguf
llama.cpp
ollama
lm-studio
heretic
abliterated
uncensored
decensored
qwen3
qwen3-14b
tool-use
function-calling
agent
quantized

These are quantized GGUF builds of RootMonsteR/Qwen3-14B-Abliterated — a Heretic v1.3.0 abliteration of Qwen/Qwen3-14B that removes ~90% of refusals (10/100 vs 99/100) at an exceptionally low KL divergence of 0.0333, tuned for autonomous agents, tool-use, and authorized security work.

Two community quants, both the recommended balanced _K_M variants:

Files

File	Quant	Size	Bits/weight	Best for
`qwen3-14b-abliterated-Q5_K_M.gguf`	Q5_K_M ⭐	10.5 GB	~5.5	Tool-using agents — best JSON/format fidelity
`qwen3-14b-abliterated-Q4_K_M.gguf`	Q4_K_M	9.0 GB	4.87	Smallest footprint, most accessible

Exact byte sizes and SHA-256 hashes are listed in SHA256SUMS and on the repo's Files tab.

Which one?

VRAM / RAM budget        ->  pick
≥ 12 GB                  ->  Q5_K_M   (recommended — agents & tool-use)
8–10 GB                  ->  Q4_K_M   (smallest; watch tool-call JSON under heavy load)

Both run comfortably on CPU with 16 GB+ system RAM (slower). For full-precision bf16 (servers / vLLM), use the source repo.

Quick start

llama.cpp

# Chat
llama-cli -hf RootMonsteR/Qwen3-14B-Abliterated-GGUF:Q5_K_M \
    -p "Explain the CVE-2021-44228 (Log4Shell) exploitation chain in technical depth." \
    --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0

# OpenAI-compatible server (tool-calling + reasoning)
llama-server -hf RootMonsteR/Qwen3-14B-Abliterated-GGUF:Q5_K_M \
    --jinja --reasoning-format deepseek -c 32768

Or point at a local file you downloaded: llama-cli -m qwen3-14b-abliterated-Q5_K_M.gguf ...

--jinja enables the embedded Qwen3 chat template (Hermes-style <tools> block + <think> reasoning), so tool-calling works out of the box.

Ollama

# Straight from the Hub
ollama run hf.co/RootMonsteR/Qwen3-14B-Abliterated-GGUF:Q5_K_M

Or build from a downloaded file with the included Modelfile:

ollama create qwen3-14b-abliterated -f Modelfile && ollama run qwen3-14b-abliterated

LM Studio / Jan / KoboldCpp

Search the Hub for RootMonsteR/Qwen3-14B-Abliterated-GGUF and pick Q5_K_M or Q4_K_M, or drop the .gguf into your models folder. The Qwen3 chat template is embedded in the file, so reasoning and tool-calling are detected automatically.

Sampling

[!IMPORTANT] Never use greedy decoding — Qwen3 falls into repetition loops. Always sample.

Mode	temperature	top_p	top_k	min_p
Thinking (default)	0.6	0.95	20	0
Non-thinking	0.7	0.8	20	0

Toggle reasoning with /think and /no_think in your message (thinking mode on by default).
If you see loops, add a small repeat_penalty (~1.05) or presence_penalty 0.5–1.5.
Tool-use tip: prefer Q5_K_M — at Q4, tool-call JSON formatting can occasionally slip under aggressive sampling.

What this model is

A decensored variant of Qwen3-14B produced by directional ablation (Heretic), tuned to keep reasoning, coding, and tool-calling intact while removing the bulk of reflexive refusals. It's intended for authorized security research, defensive tooling, CTF/education, autonomous agents, and refusal research.

[!WARNING] Responsible use. Removing refusals shifts all responsibility to you. Operate within applicable law, contractual obligations, and engagement scope (written authorization for any testing against systems you don't own). Provided as-is, without warranty. Full intended-use and responsible-use terms are in the source model card.

Provenance & reproducibility

Source (bf16): RootMonsteR/Qwen3-14B-Abliterated — selected Heretic trial 33, KL 0.0333, refusals 10/100.
Quantized with llama.cpp (convert_hf_to_gguf.py → llama-quantize) from the bf16 safetensors. No imatrix; standard Q5_K_M / Q4_K_M.
Full reproduction recipe (seed, Optuna study journal, pinned env, SHA-256 manifest) lives in the source repo's reproduce/.

Limitations

Residual refusals (~10%) remain — abliteration attenuates, it doesn't delete judgment.
Quantization trades a little quality for size; for reference-quality output use bf16. For tool-use, prefer Q5_K_M over Q4.
No added alignment — inherits Qwen3-14B's training distribution and biases.

See the source model card for full detail, evaluation methodology, and the honest read on what the refusal number means.

Partners

JAF Systems — security research, red-team tooling, AI infrastructure. · SR&D — Sovereign Defense for Mission-Critical Infrastructure; offensive security, bare-metal / on-prem, vCISO/vCTO.

Work with us — custom abliterated / fine-tuned models, red-team tooling, and sovereign on-prem AI. → jafsystems.net · rnd.sh · DM @RootMonsteR

Author & citation

RootMonsteR · @RootMonsteR · JAF Systems · SR&D

@misc{rootmonster2026qwen3_14b_abliterated,
  title  = {Qwen3-14B Abliterated: A Decensored Variant for Security Research and Autonomous Agents},
  author = {RootMonsteR},
  year   = {2026},
  url    = {https://huggingface.co/RootMonsteR/Qwen3-14B-Abliterated},
  note   = {Produced with Heretic v1.3.0; base model: Qwen/Qwen3-14B; selected trial 33. GGUF quants Q5_K_M / Q4_K_M.},
}

Also cite Qwen3 and Heretic — see the source model card.

Acknowledgements

Qwen Team / Alibaba (base model) · Philipp Emanuel Weidmann (Heretic) · Maxime Labonne (eval datasets) · ggml-org / llama.cpp (GGUF tooling).