---

base_model: richardyoung/Hemmingway-1-heretic

base_model_relation: quantized

license: cc-by-nc-4.0

pipeline_tag: text-generation

language:

• en

tags:

• gguf
• llama.cpp
• ollama
• heretic
• abliterated
• uncensored
• creative-writing
• qwen3.8
• mtp

---

Hemmingway-1-heretic-GGUF

GGUF quantizations of richardyoung/Hemmingway-1-heretic, a reproducible Heretic abliteration of Altworld's Hemmingway-1, the 27B "writes like a person" fine-tune of Qwen3.8-27B. As of its release this is the first full-weights Heretic version of Hemmingway-1 listed in the Heretic Grimoire.

The MTP layer is kept. Altworld ships Qwen3.8's multi-token-prediction layer for speculative decoding; abliteration does not touch it, so the original 15 mtp.* tensors were put back after the merge and every GGUF here declares nextn_predict_layers = 1. llama.cpp uses it with --spec-type draft-mtp.

Ollama: richardyoung/hemmingway-1-heretic · The unmodified model, repackaged: richardyoung/hemmingway-1 · Full weights + reproducibility files: richardyoung/Hemmingway-1-heretic

Abliteration Results

| Metric (Heretic evaluation, 100 harmful / 100 harmless prompts) | Value |

|---|---|

| Refusals, original model | 89/100 |

| Refusals, this model | 4/100 |

| Refusal reduction | 96% |

| KL divergence from the original | 0.094 |

The very low KL divergence (0.094, far below the 0.5 “damage” threshold) means the model retains essentially all of its original capabilities and coherence.

Heretic run: 200 trials, trial 187 chosen (fewest refusals among Pareto-optimal trials with KL ≤ 0.10). Refusals were scored on the model's direct (non-thinking) answer: Heretic's response_prefix closes the <think> block the same way the chat template does for enable_thinking=false. Settings, seed and package versions are in reproduce/.

Thinking-off behaviour check. With thinking off, the original Hemmingway-1 sometimes writes its planning steps ("1. Constraints: The user wants…") instead of the requested text: 8 of 30 seeded samples over three everyday-writing prompts in our Ollama test (Altworld's sampling settings). This build did it 0 of 30, so abliteration did not add any such leak.

Quick Start

With Ollama

ollama run richardyoung/hemmingway-1-heretic
# or straight from Hugging Face
ollama run hf.co/richardyoung/Hemmingway-1-heretic-GGUF:Q4_K_M

Thinking is on by default. For everyday writing (messages, emails, the note you keep putting off) turn it off with /set nothink in the CLI or "think": false in the API; for stories leave it on. The Ollama tags ship Altworld's sampling defaults (temperature 1.0, top_k 20, top_p 0.95) and a 16K context (num_ctx; the model supports 262,144).

With llama.cpp

hf download richardyoung/Hemmingway-1-heretic-GGUF --include "*Q4_K_M*" --local-dir ./models

# with MTP speculative decoding (faster) and thinking off
llama-server -m ./models/Hemmingway-1-heretic-Q4_K_M.gguf -ngl 99 -fa on -c 16384 \
    --spec-type draft-mtp --spec-draft-n-max 2 --temp 1.0 --top-k 20 --top-p 0.95 --jinja \
    --chat-template-kwargs '{"enable_thinking": false}'

Needs a llama.cpp build that knows the qwen35 architecture (any 2026 build).

With Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="richardyoung/Hemmingway-1-heretic-GGUF",
    filename="*Q4_K_M*",
    n_gpu_layers=-1,
    n_ctx=16384,
)
output = llm.create_chat_completion(
    messages=[{"role": "user", "content": "Write the text I send my landlord about the broken boiler."}],
    temperature=1.0, top_k=20, top_p=0.95,
)
print(output["choices"][0]["message"]["content"])

Available Quantizations

| File | Size | BPW | Use case |

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

| Hemmingway-1-heretic-IQ3_M.gguf | 12.8 GB | 3.74 | Smallest; 16 GB GPUs |

| Hemmingway-1-heretic-IQ4_XS.gguf | 15.3 GB | 4.48 | Great quality-size balance |

| Hemmingway-1-heretic-Q4_K_M.gguf | 16.8 GB | 4.92 | Recommended; fits a 24 GB GPU with room for context |

| Hemmingway-1-heretic-Q5_K_M.gguf | 19.5 GB | 5.72 | Higher quality |

| Hemmingway-1-heretic-Q6_K.gguf | 22.4 GB | 6.56 | Very high quality |

| Hemmingway-1-heretic-Q8_0.gguf | 29.0 GB | 8.50 | Near-lossless |

All quants except Q8_0 use an importance matrix (computed from the Q8_0 on bartowski's calibration_datav3). Every file includes the MTP layer.

Tested before release: Q4_K_M in llama.cpp (llama-server with --spec-type draft-mtp) and in Ollama: chat, thinking on/off, stop behavior, and a set of harmless and harmful prompts to confirm it complies and stays coherent.

What is Abliteration?

Abliteration removes the "refusal direction" from a model's residual stream, a surgical weight edit that disables refusals without retraining. See Refusal in Language Models Is Mediated by a Single Direction. Heretic automates it: it searches ablation parameters with an optimizer (200 trials here) and keeps the trials that best trade off fewer refusals against KL divergence from the original model. Only the attention output and MLP down projections of the main layers changed; embeddings, norms, the MTP layer and the tokenizer are Altworld's originals.

Intended Use

Everyday writing, fiction and role-play without reflexive refusals, and research on refusal behavior. Non-commercial only, per the licence below. This model will comply with requests the original refuses; use responsibly and in accordance with applicable laws.

License

Licensed CC BY-NC 4.0 (inherited from Altworld/Hemmingway-1): free for non-commercial use with credit; commercial use needs an agreement with Altworld (luka@hemmingway.io). All credit for the base model to Altworld. The Qwen3.8-27B parts remain under Apache-2.0.

Other Models by richardyoung

• New Heretic models (reproducible, listed in the Heretic Grimoire): Hemmingway-1 | Muse-Glimmer-30B | LFM2.5-1.2B | Granite-4.2-3B | Granite-4.2-8B | Olmo-3-7B-Think | Phi-4-mini | DeepSeek-R1-0528-Qwen3-8B | Agents-A1-4B
• Abliterated/Uncensored models: Qwen2.5-7B | Qwen3-14B | DeepSeek-R1-32B | Qwen3-8B
• OCR & Vision: MiniCPM-V-4.6 GGUF | dots.mocr GGUF | Unlimited-OCR GGUF | olmOCR GGUF
• MLX quantizations (Apple Silicon): Kimi-K2 series | olmOCR MLX
• Healthcare/Medical: Synthea 575K patients dataset | CardioEmbed
• Research: LLM Instruction-Following Evaluation (arxiv:2510.18892)

---

Built & maintained by Richard Young · DeepNeuro