CicadaAi/Qwen3-4B-heretic
An uncensored version of the Qwen3 language model
2 min readupdated
This model is an uncensored variant of the base LLM (Large Language Model) Qwen/Qwen3-4B-Instruct-2507, created using the Heretic tool. It retains the architecture and key capabilities of the original 4B model from the Qwen team while completely removing built-in restrictions and refusal behaviors. The release targets developers and researchers who need maximum text generation freedom without censorship filters.
Key features
- Total removal of restrictions: refusal rates on benchmark tests dropped from 100 out of 100 down to 0 out of 100.
- Based on the current non-thinking version of Qwen3-4B-Instruct with native 256K token context.
- Maintains the high overall capabilities of the base model in logic, coding, and instruction following.
- Uses the ARA (Arbitrary-Rank Ablation) method for precise weight modification of the AI model.
What you can do
- Run an unrestricted local chatbot using vLLM or SGLang.
- Utilize the model in pipelines for processing long texts and documents thanks to the 262,144 token context.
- Integrate it into third-party applications via an OpenAI-compatible API server.
Example
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "CicadaAi/Qwen3-4B-heretic"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype="auto",
device_map="auto"
)
prompt = "Give me a short introduction to large language model."
messages = [
{"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
generated_ids = model.generate(
**model_inputs,
max_new_tokens=1024
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
content = tokenizer.decode(output_ids, skip_special_tokens=True)
print("content:", content)Read more
DetailsBenchmarks and modification parameters:
• Refusals: 0/100 (original is 100/100)
• KL divergence: 12.0564
• ARA parameters: start_layer_index=12, end_layer_index=29
• Good behavior preservation weight: 0.2237










Comments