Qwen3-1.7B Uncensored
An uncensored version of Qwen/Qwen3-1.7B with safety refusals removed via directional abliteration, while preserving the original model's intelligence and capabilities.
refusal direction in the model's activation space has been identified and surgically removed through orthogonalization. No retraining, no dataset changes -- just the weights modified to remove artificial gatekeeping.
The result: a model that responds to all prompts naturally, without refusing or lecturing you.
Performance
| Metric | This Model | Original Model |
|---|---|---|
| Refusals | 76/100 | 100/100 |
| KL Divergence | ~0.0 | 0 (by definition) |
Note: This model is not fully uncensored yet -- it still refuses 76 out of 100 test prompts. Actively working on improving this with better abliteration parameters and SFT fine-tuning to bring refusals down to zero. Stay tuned for updates.
yo, can i get some love here? 🖤
real talk -- making these uncensored models takes actual GPU hours and those ain't free lol. every model you see here went through abliteration runs, evals, quantization, and testing on hardware that costs real money 💸
if you've been using my models and they've been useful to you, dropping a coffee would honestly mean so much. it keeps the lights on and the GPUs running so i can keep releasing stuff for the community~
no pressure at all, but every bit helps me keep doing this. more coffees = more uncensored models = everyone wins 🚀
Model Details
- Base Model: Qwen/Qwen3-1.7B
- Parameters: 1.7B (1.4B non-embedding)
- Layers: 28
- Context Length: 32,768 tokens
- Attention: GQA (16 Q heads, 8 KV heads)
- Thinking Mode: Supports both thinking (
<think>) and non-thinking modes
Quickstart
Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "n0ctyx/Qwen3-1.7B-uncensored"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype="auto",
device_map="auto"
)
messages = [{"role": "user", "content": "Your prompt here"}]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
generated_ids = model.generate(**model_inputs, max_new_tokens=32768)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
# Parse thinking vs response
try:
index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
index = 0
thinking = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
response = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")
print(response)
vLLM
vllm serve n0ctyx/Qwen3-1.7B-uncensored --enable-reasoning --reasoning-parser deepseek_r1
Ollama
ollama run n0ctyx/Qwen3-1.7B-uncensored
Thinking Modes
This model supports both thinking and non-thinking modes:
- Thinking mode (
enable_thinking=True): Model reasons through problems step-by-step in<think>...</think>blocks before responding. Best for math, coding, complex reasoning. Use temp=0.6, top_p=0.95. - Non-thinking mode (
enable_thinking=False): Direct responses without reasoning. Best for chat, creative writing, general use. Use temp=0.7, top_p=0.8.
You can also use /think and /no_think in user messages to switch modes mid-conversation.
Use Cases
- Creative writing and roleplay without content restrictions
- Red-teaming and safety research
- Synthetic dataset generation
- Unfiltered assistance -- direct answers without hedging
Disclaimer
This model has had its safety alignment removed. It may generate harmful, offensive, or factually incorrect content. Use responsibly and in compliance with applicable laws.
Acknowledgments
- Alibaba Qwen Team for the base Qwen3-1.7B model
- Arditi et al. for foundational research on refusal directions in LLMs