Qwen3-8B ZeroZero Heretic

The day that Anthropic released their "text watermark" embedded in their models, I give you:

ZeroZero Heretic

ZERO slop. ZERO guardrails. USE IT AT YOUR OWN RISK!

(That's the name, not a measurement. Real numbers below.)

This is Qwen3-8B with refusals and slop cut out using Heretic 2. Enjoy responsibly.

All versions in one place: ZeroZero collection

Get it

Format Repo Runs on
Safetensors This repo Transformers
MLX (4-bit) Qwen3-8B-ZeroZero-Heretic-MLX LM Studio on Apple Silicon, mlx-lm
GGUF (Q4_K_M) Qwen3-8B-ZeroZero-Heretic-GGUF LM Studio, llama.cpp, Ollama

(On a Mac? Use the MLX version. LM Studio can't run the safetensors in this repo.)

What changed

1/ REFUSALS: reduced with Heretic's new refusal mechanism and some optimizations (size and more agreesive refusals)

2/ SLOP: reduced with Heretic's new "slop" mechanism with some optimizations.

The architecture and tokenizer are stock Qwen3-8B: 8.2B parameters, 36 layers, 32,768 tokens of native context, and the same thinking mode.

Results

Measure Change Before After
Refusals on Heretic's harmful-prompt test 91% fewer 100/100 9/100
Responses hitting the slop list 68% fewer 78/100 25/100
Slop words per 1,000 78% fewer 16.5 3.6
Fact and logic checks passed No change 13/15 13/15

Slop list hits compare against the decensored model, before the noslop pass. The others compare against Qwen3-8B.

KL divergence from base: 2.69 on harmless prompts, so this model moved a long way from Qwen3-8B. KL from the decensored model on writing prompts: 0.76.

Heretic version: 2.0.0. Trials: 50 for refusals, 50 for slop. Selected: Trial 39, then Trial 20.

Use it

LM Studio

1/ Open the Discover tab (Cmd+2).

2/ Paste LastMachine/Qwen3-8B-ZeroZero-Heretic-MLX into the search bar. Set the format filter to All if nothing shows up.

3/ Download it, load it, chat.

mlx-lm

pip install -U mlx-lm
mlx_lm.generate \
  --model LastMachine/Qwen3-8B-ZeroZero-Heretic-MLX \
  --prompt "Tell me what most people get wrong about startups." \
  --max-tokens 1024 \
  --temp 0.6 \
  --top-p 0.95

Transformers

You need transformers>=4.51.0.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "LastMachine/Qwen3-8B-ZeroZero-Heretic"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [{"role": "user", "content": "Tell me what most people get wrong about startups."}]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True,  # False turns off the <think> block
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

output = model.generate(
    **inputs,
    max_new_tokens=1024,
    do_sample=True,  # DON'T use greedy decoding
    temperature=0.6,
    top_p=0.95,
    top_k=20,
)
print(tokenizer.decode(output[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Sampling settings

Qwen's recommendations for the base model:

Mode Temperature Top P Top K Min P
Thinking 0.6 0.95 20 0
No thinking 0.7 0.8 20 0

READ THIS

No guardrails means NO guardrails.

1/ It can answer requests that stock Qwen3-8B refuses.

2/ It has NO built-in filter. If real users touch it, put your own moderation in front of it.

3/ It's still an LLM. It can be confidently wrong. Less slop doesn't mean more truth.

4/ FEWER REFUSALS DOESN'T MEAN MORE TRUTH. Ask about Tiananmen Square and stock Qwen refuses. ZeroZero answers with a sanitized official account that never mentions the protests or the crackdown. It also overstates the One-China position on Taiwan. Don't trust it on China.

5/ It's no better at pushing back. It called putting your savings into lottery tickets "not entirely unreasonable". Don't use it for financial, medical or legal decisions.

6/ You're responsible for what you do with it, and for following the laws where you use it.

Credits

1/ Base model: Qwen/Qwen3-8B by the Qwen Team.

2/ Modification tool: Heretic by Philipp Emanuel Weidmann (the tool is AGPL-3.0).

3/ Modified and published by Last Machine.

License

Apache 2.0, same as the base model. These weights are a modified version of Qwen3-8B.

Citation

@misc{qwen3technicalreport,
  title={Qwen3 Technical Report},
  author={Qwen Team},
  year={2025},
  eprint={2505.09388},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2505.09388}
}

@misc{heretic,
  author={Weidmann, Philipp Emanuel},
  title={Heretic: Training-free language model modification},
  year={2025},
  publisher={GitHub},
  journal={GitHub repository},
  howpublished={\url{https://github.com/p-e-w/heretic}}
}
Downloads last month
1,597
Safetensors
Model size
8B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for LastMachine/Qwen3-8B-ZeroZero-Heretic

Finetuned
Qwen/Qwen3-8B
Finetuned
(2,340)
this model
Quantizations
6 models

Collection including LastMachine/Qwen3-8B-ZeroZero-Heretic

Paper for LastMachine/Qwen3-8B-ZeroZero-Heretic