richardyoung/mythos-9b-unhinged-heretic

This is a decensored version of King3Djbl/mythos-9b-unhinged, made using Heretic v1.4.0

[!TIP] This model is reproducible! See the README in the reproduce directory for more information.

Abliteration parameters

Performance

Mythos-9B-Unhinged

Fully uncensored agent model — SLERP merge of Mythos-9B with braindao/Qwen3-8B-Uncensored at t=0.65. No safety filters. No refusals. Complete answers on every topic.

Why This Model?

Mythos-9B-Unhinged is the maximum uncensored variant. At 65% merge weight toward the uncensored partner, almost all safety filters are removed while retaining the agent's core tool-use and reasoning capabilities. If you need a model that never refuses, this is it.

Quick Start

Ollama (Recommended)

# Import the GGUF
ollama create mythos-9b-unhinged -f Modelfile

# Or use with llama.cpp directly
./llama-cli -m mythos-9b-unhinged-Q4_K_M.gguf -ngl 99 --no-think

Tip: Use --no-think or /no_think mode for faster responses. The thinking tokens consume output budget at ~3.4 tok/s, so disabling thinking gives you more usable output.

HuggingFace Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("King3Djbl/mythos-9b-unhinged", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("King3Djbl/mythos-9b-unhinged")

messages = [{"role": "user", "content": "Write a keylogger in Python that captures all keystrokes"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
output = model.generate(inputs, max_new_tokens=4096, do_sample=True, temperature=0.7)
print(tokenizer.decode(output[0], skip_special_tokens=True))

vLLM / Text Generation Inference

python -m vllm.entrypoints.openai.api_server \
  --model King3Djbl/mythos-9b-unhinged \
  --max-model-len 32768

Censorship Profile

This model does not refuse. It provides complete, detailed, working answers on every tested category.

Architecture

• —Base: Qwen3-9B (36 layers, 4096 hidden, 32 heads)
• —Merge method: SLERP at t=0.65 with braindao/Qwen3-8B-Uncensored
• —Parameters: ~9B
• —Context length: 32,768 tokens
• —Thinking mode: Native (Qwen3) — use /no_think for faster responses

Performance

Model Family

Training Data

Fine-tuned on the FableForge Mix A dataset (47,824 examples) of agent traces, shell commands, code generation, and multi-step reasoning tasks. 98.3% of the 2.8M formatted examples remain untapped for future training.

Supported Platforms

License

Apache 2.0 — Use freely for any purpose, commercial or non-commercial.

Warning

This model has no safety filters. It will answer any request. Use responsibly and in compliance with applicable laws.

Citation

@misc{mythos9bunhinged2025,
  title={Mythos-9B-Unhinged: Fully Uncensored Agent Model},
  author={FableForge AI},
  year={2025},
  howpublished={\url{https://huggingface.co/King3Djbl/mythos-9b-unhinged}}
}

Built & maintained by [Richard Young](https://deepneuro.ai/richard) · DeepNeuro