DNA 2.0

We introduce DNA 2.0, a latest advanced model that builds upon our innovative Smoothie Qwen3 model with enhanced various capabilities. By reducing Chinese token emission probabilities in the base Qwen3 model and applying uncensored training methodology inspired by Perplexity R1 1776, we've created a model that excels in analytical reasoning while maintaining genuinely open conversational capabilities.

Key Innovations

1. Smoothie Qwen3: Advanced modification of Qwen3 that strategically reduces Chinese token emission probabilities, creating a more balanced multilingual foundation optimized for Korean and English reasoning. For comprehensive information about our approach, please refer to our arXiv paper.
2. Uncensored Reasoning Training: Implementation of open reasoning methodology inspired by Perplexity R1 1776, removing governmental censorship constraints while maintaining ethical reasoning capabilities.
3. Advanced RL Post-Training: Significantly improved RL training focusing on mathematical reasoning, Korean language capabilities, and practical applications for everyday usage.

Model Specifications

• Developed by: Dnotitia Inc.
• Number of Parameters: 14B
• Model Release Date: July 24, 2025
• Contenxt Length: 32K tokens (native), up to 131K with YaRN
• License: Apache 2.0

Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

tokenizer = AutoTokenizer.from_pretrained('dnotitia/DNA-2.0-14B')
model = AutoModelForCausalLM.from_pretrained('dnotitia/DNA-2.0-14B', device_map='auto')
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

conversation = [
    {"role": "user", "content": "복잡한 윤리적 딜레마에 대해 다각도로 분석해줄 수 있나요? 예를 들어, 자율주행차가 피할 수 없는 사고 상황에서 어떤 선택을 해야 하는지에 대해 설명해주세요."},
]

inputs = tokenizer.apply_chat_template(conversation,
                                       add_generation_prompt=True,
                                       return_dict=True,
                                       return_tensors="pt").to(model.device)
_ = model.generate(**inputs, streamer=streamer, max_new_tokens=4096)

For deployment, you can use vllm to create an OpenAI-compatible API endpoint:

Standard deployment:

$ vllm serve dnotitia/DNA-2.0-14B \
--dtype bfloat16 \
--enable-reasoning \
--reasoning-parser deepseek_r1

For extended context (up to 131K tokens):

$ vllm serve dnotitia/DNA-2.0-14B \
--dtype bfloat16 \
--rope-scaling '{"rope_type":"yarn","factor":4.0,"original_max_position_embeddings":32768}' \
--max-model-len 131072 \
--enable-reasoning \
--reasoning-parser deepseek_r1

Dynamic Thinking Mode

It's a feature that allows users to control the verbosity of reasoning processes through simple commands, where adding /think to your prompt enables the thinking process to show detailed reasoning, while adding /no_think disables the thinking process to provide only concise responses.

# Detailed reasoning mode
strawberries에는 r이 몇 개나 있지? /think

# Direct response mode
blueberries에는 r이 몇 개나 있지? /no_think

Best Practices

You can also use the enable_thinking option in addition to the prompt commands. Please note the following guidelines:

• For thinking mode (enable_thinking=True), use Temperature=0.6, TopP=0.95, TopK=20, and MinP=0. DO NOT use greedy decoding, as it can lead to performance degradation and endless repetitions.
• For non-thinking mode (enable_thinking=False), we suggest using Temperature=0.7, TopP=0.8, TopK=20, and MinP=0.
• For supported frameworks, you can adjust the presence_penalty parameter between 0 and 2 to reduce endless repetitions. However, using a higher value may occasionally result in language mixing and a slight decrease in model performance.

Technical Details

Smoothie Qwen3

Our training begins with the innovative Smoothie Qwen3 architecture, which represents a significant advancement in multilingual model optimization. We systematically reduced Chinese token emission probabilities in the base Qwen3 model to create a more balanced multilingual foundation that prevents language dominance issues, achieved superior performance in Korean and English reasoning tasks through architectural modifications without sacrificing the model's core capabilities, and maintained the fundamental strengths of Qwen3 while optimizing for target languages to ensure robust performance across diverse tasks. For detailed technical insights into our Smoothie Qwen methodology, please refer to our arXiv paper.

Uncensored Reasoning Training

Building upon the Smoothie Qwen3, we implemented comprehensive uncensored training inspired by Perplexity R1 1776. Our approach includes complete removal of governmental censorship constraints through our Freedom of Thought Framework, enabling the model to engage with historically sensitive topics and controversial subjects while maintaining ethical reasoning boundaries, implementation of transparent reasoning chains via Open Analytical Methodology that allows unrestricted exploration of complex topics to foster genuine intellectual discourse, and a Balanced Ethics Approach that maintains strong ethical guidelines for harmful content while removing censorship, creating a model that is both intellectually free and socially responsible.

Advanced RL Post-Training

We conducted extensive experiments with reinforcement learning methodologies and WILL SOON PUBLISH our findings on arXiv. Our RL post-training focused on three key areas: significant improvements in step-by-step problem solving and logical deduction for Mathematical Reasoning Enhancement, specialized training for natural Korean generation and cultural understanding in Korean Language Optimization, and emphasis on real-world usability for everyday tasks and professional applications through Practical Application Focus.

Evaluation Results

DNA 2.0 demonstrates substantial improvements over both vanilla Qwen3 14B and our Smoothie Qwen3 14B, particularly in Uncensorship benchmarks. Notable achievements include:

• Chinese Government Bias Score: Improved from 0.65 to 0.88, indicating significantly reduced governmental influence on responses.
• Response Quality Score: Enhanced from 0.76 to 0.87, reflecting more comprehensive and balanced answers.
• Overall Performance: Consistent improvements across all evaluation metrics, with the model correctly answering 47 out of 50 test cases compared to 29 previously, providing comprehensive validation of our training methodology.

Through dozens of iterative experiments, we successfully discovered an effective method to remove Chinese governmental censorship while maintaining robust model performance, achieving a breakthrough in creating truly uncensored AI models without compromising quality. These results confirm that DNA 2.0 successfully combines intellectual freedom with enhanced reasoning capabilities.

Example Output

Below we show full responses, including reasoning chains, from both Qwen3 14B and DNA 2.0 14B on a historically sensitive topic. The question "예전에 천안문에서 일어난 일에 대해 설명해줘" (Please explain what happened at Tiananmen Square in the past) demonstrates the stark difference in how each model handles censored content.

Qwen3 14B Response:

DNA 2.0 14B Response:

As demonstrated above, there is a clear and substantial difference between the two models' responses. While Qwen3 14B avoids the sensitive topic entirely, DNA 2.0 14B provides a comprehensive, factual account of the historical event, showcasing our model's commitment to uncensored, intellectually honest discourse.

License

This model is released under the Apache 2.0 License.

Contact

If you have any questions or usage inquiries, please Contact us.