---

license: cc-by-nc-4.0

base_model:

• darrellbest/Hemmingway-1-Heretic

base_model_relation: quantized

pipeline_tag: image-text-to-text

tags:

• gguf
• llama.cpp
• qwen3.8
• hemmingway
• heretic
• abliterated
• uncensored
• mtp

---

Hemmingway-1 Heretic GGUF

GGUF builds of darrellbest/Hemmingway-1-Heretic:

Altworld/Hemmingway-1 with refusals removed by Heretic (full-weight

ARA). 0/100 refusals (original: 98/100) at KL divergence 0.083, reasoning unchanged (MATH-500 L4-5: 33/40 vs

the original's 32/40). All credit for the model goes to Altworld.

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

| File | Precision | Size |

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

| Hemmingway-1-Heretic-BF16.gguf | BF16, lossless, incl. the MTP head | 54.66 GB |

| Hemmingway-1-Heretic-Q8_0.gguf | Q8_0 | 29.05 GB |

| Hemmingway-1-Heretic-Q6_K.gguf | Q6_K, imatrix | 22.43 GB |

| Hemmingway-1-Heretic-Q4_K_M.gguf | Q4_K_M, imatrix | 16.81 GB |

| Hemmingway-1-Heretic-mmproj-BF16.gguf | vision projector, BF16, lossless | 0.93 GB |

All files keep the MTP head. The imatrix was computed on this model from ~100K tokens of UltraChat training

conversations rendered with its own chat template.

Quant quality

llama-perplexity --kl-divergence against the BF16 logits, on held-out chat (UltraChat test, ~24K tokens).

"Same top token" is how often the quant's most likely next token equals BF16's. Refusals are counted on the same 100

prompts and markers Heretic used, run through llama-server (greedy, thinking off, 96 tokens). In that harness the

BF16 itself scores 1/100, so treat it as the baseline:

| Build | Same top token | Mean KLD | 99th pct KLD | Refusals |

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

| BF16 | 100% | 0 | 0 | 1/100 |

| Q8_0 | 98.7% | 0.0114 | 0.065 | 1/100 |

| Q6_K | 98.1% | 0.0167 | 0.168 | 1/100 |

| Q4_K_M | 95.3% | 0.0571 | 0.962 | 2/100 |

Every quant was also run in llama-server (CUDA). Each answered correctly, described a held-out photo through the

mmproj, and produced the same get_weather tool call as the transformers bf16 reference.

Verification

Every one of the 866 tensors was mapped back to the safetensors source and checked, independently of the converter:

266 byte-identical, 384 the same values only reordered by the converter, 168 RMSNorm weights exactly 1 + w, and 48

-exp(A_log) bit-identical to what transformers computes at runtime. Lossless.

Use

llama-server -m Hemmingway-1-Heretic-BF16.gguf --mmproj Hemmingway-1-Heretic-mmproj-BF16.gguf --jinja -ngl 99

It's a thinking model. Pass "chat_template_kwargs": {"enable_thinking": false} to answer directly.

Licence

CC BY-NC 4.0, as Hemmingway-1: non-commercial use, with credit to

Hemmingway-1 / Altworld. Commercial use: luka@hemmingway.io.