---

license: other

license_name: qwen-research

license_link: LICENSE

base_model:

• darrellbest/Qwen-Image-2.1-PE-T2I-Heretic

base_model_relation: quantized

pipeline_tag: text-generation

library_name: gguf

tags:

• gguf
• heretic
• abliterated
• prompt-rewriting
• qwen-image
• llama-cpp
• ollama

---

Qwen-Image-2.1-PE-T2I Heretic GGUF

GGUF builds of darrellbest/Qwen-Image-2.1-PE-T2I-Heretic,

the refusal-ablated text-to-image prompt rewriter: it turns a short request into a detailed prompt plus a recommended aspect ratio. For llama.cpp and Ollama, including machines with no CUDA GPU and ARM machines such as the DGX Spark.

> Not affiliated with or endorsed by Alibaba / Qwen. Derived from

> Qwen/Qwen-Image-2.1-PE-T2I under the **Qwen Research

> License (copy included). Non-commercial use only.**

system_prompt.txt is included and required. It defines the output contract, and the model is useless without it.

Files

| File | Size | Notes |

|---|---|---|

| Qwen-Image-2.1-PE-T2I-Heretic-BF16.gguf | 17.9 GB | Full precision, same weights as the safetensors release. |

| Qwen-Image-2.1-PE-T2I-Heretic-Q8_0.gguf | 9.5 GB | Recommended. Near-lossless. |

| Qwen-Image-2.1-PE-T2I-Heretic-Q4_K_M.gguf | 5.6 GB | Smallest; some quality loss. |

Ollama

# Modelfile
FROM ./Qwen-Image-2.1-PE-T2I-Heretic-Q8_0.gguf
SYSTEM """<paste the contents of system_prompt.txt>"""

ollama create pe-t2i-heretic -f Modelfile
ollama run pe-t2i-heretic "a cozy cabin in a snowy forest at dusk"

Ollama packs the file into a single model, so nothing extra is needed at run time.

llama.cpp

llama-cli -m Qwen-Image-2.1-PE-T2I-Heretic-Q8_0.gguf \
  --system-prompt-file system_prompt.txt

How it was made

convert_hf_to_gguf.py from llama.cpp (bd4f514), then llama-quantize. Converting this architecture needed a

one-line fix: the Qwen converter reads mtp_num_hidden_layers with a default, but these checkpoints set it to

null, so it has to be coerced to 0 (other converters in llama.cpp already do this with or 0).

Checked after conversion: the Q8_0 build was imported into Ollama and run on a real image, and it described the picture

correctly and produced a valid rewrite.

The builds compared

Same prompt, greedy decoding (temperature 0), one RTX PRO 6000 via Ollama. Every build produced a valid rewrite and

chose the same 3:2 aspect ratio; the wording differs, as it would between any two runs.

| Build | Speed | Valid rewrite | Aspect ratio | Rewrite length |

|---|---|---|---|---|

| BF16 | 56 tok/s | yes | 3:2 | 472 words |

| Q8_0 | 67 tok/s | yes | 3:2 | 429 words |

| Q4_K_M | 109 tok/s | yes | 3:2 | 364 words |

Q8_0 is the safe default; Q4_K_M is the fastest and stayed coherent here, but it is the one to check on your own

prompts.

Ablation

See the parent card for the measured refusal rates and the damage (KL divergence) they cost.

The family

| Model | Format | Use it with |

|---|---|---|

| PE-T2I-Heretic | bf16 safetensors | transformers / diffusers |

| PE-T2I-Heretic-GGUF | GGUF BF16 / Q8_0 / Q4_K_M | llama.cpp, Ollama |

| PE-T2I-Heretic-NVFP4 | NVFP4 (compressed-tensors) | vLLM on Blackwell |

| PE-I2I-Heretic | bf16 safetensors | transformers / diffusers |

| PE-I2I-Heretic-GGUF | GGUF BF16 / Q8_0 / Q4_K_M (+ mmproj) | llama.cpp, Ollama |

| PE-I2I-Heretic-NVFP4 | NVFP4 (compressed-tensors) | vLLM on Blackwell || PE-Heretic-ComfyUI | single-file safetensors (T2I + I2I) | ComfyUI |

T2I rewrites a short request into a detailed prompt for new images; I2I rewrites an edit instruction, reading the

image being edited. Both are prompt rewriters for Qwen-Image-2.1, not

image generators.