Qwen3-8B-heretic
an uncensored version of Qwen3-8B
2 min readupdated
Developer RProrochenko has released an uncensored modification of the Qwen/Qwen3-8B language model (8.19B parameters, 36 layers), created using Heretic v2.1.0 via Arbitrary-Rank Ablation (ARA — targeted removal of refusal directions in the model's weights). In practice, the model responds to prompts that the original Qwen3-8B declines: the refusal rate drops from 99/100 to 5/100 with minimal deviation from the original distribution (KL divergence 0.0267). At the same time, it retains all core capabilities of the Qwen3 family: reasoning in thinking mode, coding, tool calling, and support for over 100 languages. It is suited for developers and hobbyists who need a compact local LLM without strict filters for roleplay, creative writing, and research.
Key features
- Sharp drop in refusals: the Refusals score falls from 99/100 on the original Qwen3-8B down to 5/100.
- Preserved baseline quality: a low KL divergence (0.0267) shows that the modification only targets the refusal mechanism across layers 0–35, leaving the rest of the model's behavior largely unchanged.
- Dual generation modes: supports seamless switching between reasoning mode (
enable_thinking=Truewith/thinktags) and fast dialogue (/no_think). - Agentic features and multilingual support: compatible with MCP and Qwen-Agent for external tool calling, alongside support for 100+ languages and dialects.
What you can do
- Run it locally in Ollama, LM Studio, llama.cpp, or MLX-LM for unrestricted brainstorming, roleplay, and text generation without triggering built-in filters.
- Serve it via vLLM or SGLang as an OpenAI-compatible API server for local chatbots and agents.
- Pair it with Qwen-Agent and MCP servers to automate tasks requiring tool calling and logical analysis.
Example
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "RProrochenko/Qwen3-8B-heretic"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
text = tokenizer.apply_chat_template([{"role": "user", "content": "Привет!"}], tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
print(tokenizer.decode(model.generate(**inputs, max_new_tokens=512)[0]))
Read more
DetailsMetrics: Refusals — 5/100 (vs. 99/100 on the base model), KL divergence — 0.0267.
Context: 32,768 tokens natively and up to 131,072 tokens with YaRN scaling.
Software: requires
transformers>=4.51.0 (older versions raise KeyError: 'qwen3').Recommendations: for thinking mode, use Temperature=0.6, TopP=0.95, TopK=20, and avoid greedy decoding.










Comments