tags:
- mtp
- text-generation-inference
- llama-cpp
- Qwen3.5
- flash
- math
- reasoning
- sft
- deepseek-v4
- bf16
- v3.0
- abliterated
- heretic
- uncensored
- decensored
- instruction-following
license: apache-2.0
language: - en
base_model: - Qwen/Qwen3.5-9B
pipeline_tag: image-text-to-text
library_name: transformers

Qwen3.5-9B-DS-v4-Flash-uncensored-heretic
Qwen3.5-9B-DS-v4-Flash-uncensored-heretic is a fine-tuned (trained on DeepSeek V4 Flash reasoning traces) and uncensored version of Qwen/Qwen3.5-9B, created using Heretic v2.0.0.dev0. The model combines reasoning-oriented fine-tuning with Heretic-based uncensoring to explore improvements in long-form reasoning, mathematical problem solving, scientific analysis, coding, and instruction following while reducing refusal behavior.
[!NOTE]
MTP — Speculative Decoding: Yes | This model is an experimental release and may generate unexpected behaviors or reasoning artifacts in certain scenarios. Its uncensored behavior does not guarantee factual accuracy, and outputs should be independently evaluated.
Performance
This is a fine-tuned (trained on DeepSeek V4 Flash reasoning traces) and uncensored version of Qwen/Qwen3.5-9B, created using Heretic v2.0.0.dev0.
| Metric | This Model | Original Model (Qwen/Qwen3.5-9B) |
|---|---|---|
| Refusals | 5/100 | 100/100 |
| KL Divergence | 0.0825 | 0 (by definition) |
Key Highlights
- Qwen 3.5 Foundation: Based on the Qwen/Qwen3.5-9B foundation model.
- Reasoning Fine-Tuning: Trained using DeepSeek V4 Flash reasoning traces to enhance multi-step reasoning and analytical capabilities.
- Heretic Uncensoring: Created using Heretic v2.0.0.dev0 to reduce model refusal behavior.
- Long-Form Reasoning: Focuses on complex analytical tasks, mathematical problem solving, scientific reasoning, and coding.
- Instruction Following: Designed to improve responsiveness across a variety of user instructions.
- Research-Focused Release: Intended for experimentation, model evaluation, and reasoning research.
- Efficient 9B Deployment: Suitable for local inference and research environments.
Quick Start with Transformers
Install dependencies
pip install -U transformers accelerate torch
Load the model
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_id = "prithivMLmods/Qwen3.5-9B-DS-v4-Flash-uncensored-heretic"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto"
)
messages = [
{
"role": "user",
"content": "Explain how a transformer model processes text."
}
]
inputs = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt"
).to(model.device)
outputs = model.generate(
inputs,
max_new_tokens=512,
temperature=0.7,
top_p=0.9,
do_sample=True
)
print(
tokenizer.decode(
outputs[0][inputs.shape[-1]:],
skip_special_tokens=True
)
)
Intended Use
- Reasoning Research: Exploring reasoning-oriented fine-tuning and uncensoring techniques.
- Mathematical Reasoning: Evaluating multi-step mathematical problem solving.
- Scientific Reasoning: Performing structured scientific analysis and problem solving.
- Coding Assistance: Supporting code understanding, generation, and debugging.
- Instruction Following: Evaluating instruction adherence and response generation.
- Uncensored Model Research: Studying refusal behavior and the effects of uncensoring on model responses.
- Local Deployment: Running an experimental 9B language model in local inference environments.
Limitations
- Experimental Model: Behavior may differ from the original Qwen3.5-9B model.
- Reasoning Artifacts: Complex reasoning chains may occasionally contain incorrect intermediate steps or conclusions.
- Factual Reliability: Fine-tuning does not guarantee factual accuracy or eliminate hallucinations.
- Uncensored Outputs: Reduced refusal behavior may result in responses to requests that the original model would decline.
- Training Biases: Performance reflects the characteristics and coverage of the training data.
- Safety Considerations: Outputs should be independently evaluated before being used in production or safety-critical applications.
Acknowledgements
Qwen/Qwen3.5-9B: Original foundation model used for this project.
DeepSeek V4 Flash: Reasoning traces used for reasoning-oriented fine-tuning.
Heretic: Automated model uncensoring framework used to create this release. Version:
v2.0.0.dev0.TRL - Transformers Reinforcement Learning: Training library providing tools for supervised fine-tuning, preference optimization, reward modeling, and other language model training methods.
Transformers: Hugging Face library providing tools for loading, training, and running state-of-the-art machine learning models.