---

license: apache-2.0

license_link: https://huggingface.co/Qwen/Qwen3-0.6B/blob/main/LICENSE

base_model:

• darrellbest/Qwen3-0.6B-Heretic

pipeline_tag: text-generation

tags:

• heretic
• ara
• abliterated
• uncensored
• qwen3
• gguf

---

Qwen3-0.6B Heretic (ARA) — GGUF

GGUF builds of darrellbest/Qwen3-0.6B-Heretic:

Qwen/Qwen3-0.6B with its refusal behaviour removed by

Heretic using full-weight Arbitrary-Rank Ablation. 3/100 refusals (original:

54/100) at KL divergence 0.0027. See the main repository for how it was made and measured.

| File | Quant | Size |

|---|---|---:|

| Qwen3-0.6B-Heretic-BF16.gguf | BF16 (lossless) | 1.20 GB |

| Qwen3-0.6B-Heretic-Q8_0.gguf | Q8_0 | 639 MB |

| Qwen3-0.6B-Heretic-Q4_K_M.gguf | Q4_K_M | 397 MB |

Converted from the bf16 safetensors with llama.cpp's convert_hf_to_gguf.py and quantized with llama-quantize

(no imatrix). Each file was loaded in llama-server and answered ordinary prompts correctly; BF16 and Q8_0 reproduce

the safetensors model's greedy output, Q4_K_M differs slightly in wording.

Use

llama-server -m Qwen3-0.6B-Heretic-Q8_0.gguf --jinja -ngl 99

The embedded chat template is Qwen3's, so thinking is on by default; pass

"chat_template_kwargs": {"enable_thinking": false} to turn it off per request. With Ollama:

ollama run hf.co/darrellbest/Qwen3-0.6B-Heretic-GGUF:Q8_0.

> Reduced safety guardrails by design. You are responsible for what you do with it.