---
license: apache-2.0
language:
- en
- zh
base_model:
- deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
tags:
- Reasoning
- Instruct
- Uncensored
- Distilled
- GGUF
- Quantized
---

# DeepSeek-R1-Distill-Qwen-7B-Uncensored

This repository hosts **uncensored and efficiency-focused builds** of **DeepSeek-R1-Distill-Qwen-7B**, intended for users who require **direct model behavior**, **strong reasoning**, and **full local control** without aggressive automated filtering.

The model is suitable for advanced experimentation, private deployments, and research scenarios where transparency and flexibility are prioritized.

---

## Model Overview

- **Model Name**: DeepSeek-R1-Distill-Qwen-7B-Uncensored  
- **Base Model**: DeepSeek-R1-Distill-Qwen-7B  
- **Architecture**: Decoder-only Transformer  
- **Parameter Count**: ~7B  
- **Modalities**: Text  
- **Context Length**: Up to 32K tokens (runtime dependent)  
- **Developer (Base)**: DeepSeek AI  
- **Distillation Target**: Qwen-based reasoning model  
- **License**: Apache-2.0 (inherits base model license)  
- **Languages**: Multilingual (English, Chinese, others)

---

## Project Intent

This release is designed for **users who want minimal behavioral constraints** while preserving the structured reasoning and instruction-following strengths of the DeepSeek-R1 distillation.

Key objectives include:

- Predictable, direct responses without heavy content suppression  
- Strong multi-step reasoning and analytical depth  
- Compatibility with local and offline inference setups  
- A solid foundation for further alignment, fine-tuning, or research  

This is **not** a consumer-safety-aligned assistant and is intended for controlled environments.

---

## Quantized Variants (GGUF)

To support a wide range of hardware, multiple GGUF quantization levels are provided.

### Q2_K (2-bit)
- Extremely small memory footprint  
- Intended for experimentation or extreme hardware constraints  
- Severe degradation in reasoning and instruction accuracy  

### Q3_K_M (3-bit)
- Slight improvement over 2-bit  
- Lightweight and fast  
- Limited suitability for complex reasoning tasks  

### Q4_K_M (4-bit)
- Strong efficiency-to-quality tradeoff  
- Works well on CPUs and low-VRAM GPUs  
- Suitable for general chat and exploratory reasoning  

### Q5_K_M (5-bit)
- Recommended default for most users  
- Retains most reasoning and instruction-following ability  
- Balanced memory usage and output quality  

### Q6_K (6-bit)
- Higher reasoning fidelity  
- Increased memory requirements  
- Better performance on long or complex prompts  

### Q8_0 (8-bit)
- Near full-precision behavior  
- Highest quality quantized variant  
- Best choice when memory is not a limiting factor  

> Output quality depends heavily on context length, sampling parameters, and inference backend.

---

## Prompting Format

The model performs best with a structured chat format:

```

<|system|>
High-level instructions or behavioral guidance
<|user|>
User prompt
<|assistant|>

````

Clear system messages are recommended to guide tone, verbosity, and task focus.

---


### Suggested Settings

* Temperature: `0.6 – 0.8` for analytical tasks
* Use `Q5_K_M` or higher for reasoning-heavy prompts
* Avoid ultra-low-bit quantizations for long-context analysis

---

## Capabilities

* Strong logical and mathematical reasoning
* Effective multi-step analysis and planning
* Clear instruction-following behavior
* Suitable for research into reasoning and alignment
* Performs well in uncensored local deployments
* Maintains coherence over extended conversations

---

## Recommended Use Cases

* Local reasoning assistants
* Research and alignment studies
* Offline analysis and experimentation
* Advanced prompt engineering workflows
* Private deployments requiring full user control

---

## Important Notes

* This model intentionally avoids strong automated moderation
* Users are responsible for ensuring lawful and ethical usage
* Not recommended for unsupervised or public-facing applications
* Quantized variants may hallucinate more than full-precision models

Always evaluate outputs in the context of your intended application.

---

## Acknowledgements

* DeepSeek AI for releasing the DeepSeek-R1 model family
* Qwen team for the underlying architecture contributions
* The `llama.cpp` and GGUF ecosystem for enabling efficient local inference
* Open-source contributors supporting transparent LLM research

---

## Contact

For issues related to quantization files or repository content, please open an issue in this repository.