• en
• zh

• deepseek-ai/DeepSeek-R1-Distill-Qwen-7B

• 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

Q3KM (3-bit)

• Slight improvement over 2-bit
• Lightweight and fast
• Limited suitability for complex reasoning tasks

Q4KM (4-bit)

• Strong efficiency-to-quality tradeoff
• Works well on CPUs and low-VRAM GPUs
• Suitable for general chat and exploratory reasoning

Q5KM (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|>

• Temperature: 0.6 – 0.8 for analytical tasks
• Use Q5KM or higher for reasoning-heavy prompts
• Avoid ultra-low-bit quantizations for long-context analysis

• 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.