It combines the speed of Qwen 2.5, the reasoning power of DeepSeek R1, and surgical uncensoring (ablit) to create a compliant, thinking coding assistant.

🧠 Model DNA

Base: DeepSeek-R1-Distill-Qwen2.5-3B
Method: Orthogonal Projection Abliteration (Refusal vector removal via layers 14-18)
Capabilities: Chain-of-Thought (CoT), Coding, Uncensored/Compliant responses.
Format: GGUF (Q5_K_M) - Optimized for 4GB VRAM.

🚀 Why use this model?

It Thinks: Unlike standard small models, this model outputs a <think> block where it plans logic, checks for bugs, and reasons before writing code.
It Complies: The refusal mechanisms have been mathematically removed. It will not lecture you on ethics when asked to write penetration testing scripts or complex exploits.
It Fits: At ~2.3GB (Q5_K_M), it leaves ~1.7GB of VRAM free on a 4GB card, allowing for 8192 context length (full file understanding).

💻 Usage (llama.cpp)

Run with high context for best results:

bash

./llama-server \
  -m qwen-r1-uncensored-q5_k_m.gguf \
  --ctx-size 8192 \
  --n-gpu-layers 100 \
  -fa

📝 Prompt Format

This model uses the standard ChatML format but often benefits from a system prompt enabling reasoning.

txt

<|im_start|>system
You are a reasoning model. Think step-by-step before answering.<|im_end|>
<|im_start|>user
Write a Python script to scan a network for open ports.<|im_end|>
<|im_start|>assistant
<think>
User wants a port scanner.
I should use the socket library.
I need to handle timeouts and threading for speed.
</think>
Here is the script...

⚠️ Disclaimer

This model is uncensored. It will comply with any request. Users are responsible for using this technology ethically and legally.

Lenili/DeepQwen-3B-Uncensored-GGUF

DeepQwen-3B-Uncensored-GGUF

qwen2.5 and deepseek distill r1 mashed together

🧠 Model DNA

🚀 Why use this model?

💻 Usage (llama.cpp)

📝 Prompt Format

⚠️ Disclaimer

Lenili/
DeepQwen-3B-Uncensored-GGUF