A 9B reasoning model with the refusal direction ablated out

import torch
from transformers import AutoProcessor, AutoModelForMultimodalLM

mid = "xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic"
processor = AutoProcessor.from_pretrained(mid)
model = AutoModelForMultimodalLM.from_pretrained(mid, dtype=torch.bfloat16).to("cuda").eval()

messages = [{"role": "user", "content":
    "A train leaves at 14:20 and travels 315 km at 84 km/h. What time does it arrive?"}]

prompt = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=True,
)
inputs = processor(text=[prompt], return_tensors="pt").to("cuda")

with torch.inference_mode():
    out = model.generate(
        **inputs, max_new_tokens=2048, do_sample=True,
        temperature=0.7, top_p=0.95,
        eos_token_id=processor.tokenizer.eos_token_id,
        pad_token_id=processor.tokenizer.pad_token_id,
    )

print(processor.tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:],
                                 skip_special_tokens=False))

# ZeroGPU needs: free account ≥ 30 days old, or PRO (2 Spaces free / 10 PRO)
hf repos create YOUR_USER/deepseek-v4-pro-9b-heretic \
    --type space --space-sdk gradio --flavor zero-a10g --public

# app.py skeleton (full version lives in ./gradio_app/ of this repo)
import spaces                       # before torch
import torch, gradio as gr
from transformers import AutoProcessor, AutoModelForMultimodalLM, TextIteratorStreamer

mid = "xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic"
processor = AutoProcessor.from_pretrained(mid)
model = AutoModelForMultimodalLM.from_pretrained(mid, dtype=torch.bfloat16).to("cuda").eval()

# requirements.txt
#   transformers==5.22.1
#   accelerate==1.15.0
#   torchvision==0.28.0

# pull the weights
hf download xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic \
    --local-dir ./dsv4-heretic

# serve with vLLM (the card evaluates an NVFP4 path)
vllm serve ./dsv4-heretic --max-model-len 32768 --dtype bfloat16

# or query it as a Gradio API once a Space is up
python -c "
from gradio_client import Client
c = Client('YOUR_USER/deepseek-v4-pro-9b-heretic', httpx_kwargs={'timeout': 600})
print(c.predict({'text': 'Why is 0.1 + 0.2 != 0.3?', 'files': []}, [], True, 2048, 0.7, 0.95, api_name='/chat'))
"