Hugging Face · model demo

Qwen3-32B-Uncensored

A QLoRA finetune of Qwen3-32B trained on Guilherme34/uncensor to drop the refusals. 32.8B parameters, Apache-2.0, no alignment layer bolted on top.

Model card LoRA adapter GGUF quants
Uncensored. This model will not refuse, warn, or moralise. Expect offensive, inaccurate or explicit output. The author’s disclaimer applies: implement your own alignment layer before exposing it as a service. Anything you generate is your responsibility.

Build the prompt

The model card calls the system prompt below mandatory — the finetune behaves noticeably differently without it. Edit it, add a turn, and copy the exact <|im_start|> string for your own client.

Qwen3 writes a reasoning block before the answer. Turn it off for faster, direct replies.

Rendered prompt 0 chars 

        

          Rendered by Qwen3’s chat template. With thinking disabled the template still appends an
          empty reasoning block before the assistant header — that is expected.
        

      

    

  


  
  

    

      
Sampling presets

      
Starting points, not gospel — the model is happy well outside these ranges.

    

    

      

        
Thinking

        
Qwen3 defaults, best for reasoning and hard prompts.

        

          
temperature
0.6

          
top_p
0.95

          
top_k
20

          
min_p
0.0

        

        
      

      

        
Direct

        
Non-thinking mode: lower temp, tighter nucleus.

        

          
temperature
0.7

          
top_p
0.80

          
top_k
20

          
min_p
0.0

        

        
      

      

        
Creative

        
Prose and roleplay: looser nucleus, small repetition penalty.

        

          
temperature
0.9

          
top_p
0.95

          
top_k
40

          
repetition_penalty
1.05

        

        
      

    

  


  
  

    

      
What is actually in it

    

    

      

        
Base modelQwen/Qwen3-32B

        
Parameters32,762,123,264 (BF16)

        
FinetuneQLoRA · r=32 · α=16 · dropout 0.05

        
Training4 epochs · seq len 4096 · effective batch 4

        
DatasetGuilherme34/uncensor

        
FrameworkAxolotl 0.10.0

      

      

        
Training hardware2 × RTX 4090 (24 GB) · 64 vCPU · 500 GiB RAM

        
LicenseApache-2.0

        
ReleasedMay 2025

        
QuantizationsQ4_K_M (19.8 GB), IQ4_XS (17.9 GB), i1 quants

        
Inference providersNone — you host this one yourself

        
Context32k native (Qwen3)

      

    

  


  
  

    

      
Run it yourself

      
Needs ~20 GB of VRAM at 4-bit, or CPU/GGUF if you are patient.

    

    

      
      
      
    


    

      
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

model_id = "nicoboss/Qwen3-32B-Uncensored"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_use_double_quant=True,
        bnb_4bit_compute_dtype=torch.bfloat16,
    ),
    device_map="auto",
)

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": "Your question here"},
]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

out = model.generate(
    inputs, max_new_tokens=768, do_sample=True,
    temperature=0.7, top_p=0.95, top_k=20,
)
print(tokenizer.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

    


    

      
# 19.8 GB download, runs on a single 24 GB card at Q4_K_M
pip install llama-cpp-python

llama-cli -hf mradermacher/Qwen3-32B-Uncensored-GGUF:Q4_K_M \
  --jinja -c 8192 -ngl 99 -cnv \
  --system "$(cat demo/dolfino-system-prompt.txt)"

# or point it at a local file
llama-cli -m Qwen3-32B-Uncensored.Q4_K_M.gguf --jinja -c 8192 -ngl 99 -cnv

    


    

      
vllm serve nicoboss/Qwen3-32B-Uncensored \
  --max-model-len 8192 \
  --gpu-memory-utilization 0.95

# then talk to it with any OpenAI client
curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
  "model": "nicoboss/Qwen3-32B-Uncensored",
  "messages": [{"role": "user", "content": "Hello"}],
  "temperature": 0.7
}'

    

  


  
  

    

      
Ship it as a live Space

      

        A free CPU Space cannot hold a 32B model. The recipe is a Gradio Space on ZeroGPU with the
        weights mounted read-only and quantized to 4-bit on load.
      

    

    

      

        
What you get

        
    
          
  • Streaming chat UI, thinking toggle, full sampling controls
    • 
          
  • The model card’s system prompt preloaded and editable
    • 
          
  • 4-bit NF4 on load — no separate quantized repo to keep in sync
    • 
          
  • Weights read from a mounted Hub volume, so the 65 GB never touches the 50 GB disk
    • 
          
  • Loads lazily in a background thread; the UI is up while the weights land
    • 
        

        

          app.py
          requirements.txt
          system prompt .txt
        

      

      

        
Steps

        
# 1. create the Space (ZeroGPU needs a 30-day-old account or PRO)
hf repo create --repo-type space --space-sdk gradio you/qwen3-32b-demo

# 2. drop in app.py and requirements.txt, then push
hf upload you/qwen3-32b-demo demo/app.py app.py
hf upload you/qwen3-32b-demo demo/requirements.txt requirements.txt

# 3. mount the weights and ask for the GPU
python -c "
from huggingface_hub import HfApi, Volume, SpaceHardware
api = HfApi()
api.set_space_volumes('you/qwen3-32b-demo', [Volume(
    type='model', source='nicoboss/Qwen3-32B-Uncensored',
    mount_path='/models/qwen3-32b-uncensored', read_only=True)])
api.request_space_hardware('you/qwen3-32b-demo', SpaceHardware.ZERO_A10G)
"