dnotitia/DNA-2.0-14B
DNA 2.0 is a 14 billion parameter language model developed by Dnotitia Inc., built upon an advanced modification of Qwen3. It features strategically reduced Chinese token emission probabilities for balanced multilingual performance, particularly excelling in Korean and English reasoning. This model is uniquely trained with an uncensored methodology, inspired by Perplexity R1 1776, to provide open conversational capabilities and enhanced analytical reasoning, supporting a 32K native context length and up to 131K with YaRN.
Loading preview...
DNA 2.0: Uncensored Reasoning and Multilingual Optimization
DNA 2.0 is a 14 billion parameter model developed by Dnotitia Inc., building on an advanced modification of the Qwen3 architecture. It is designed to excel in analytical reasoning and open conversational capabilities, particularly in Korean and English, while maintaining ethical reasoning boundaries.
Key Innovations & Capabilities
- Smoothie Qwen3 Architecture: This foundational innovation strategically reduces Chinese token emission probabilities in the base Qwen3 model, creating a more balanced multilingual foundation optimized for Korean and English reasoning. This approach is detailed in their arXiv paper.
- Uncensored Reasoning Training: Inspired by Perplexity R1 1776, DNA 2.0 implements an open reasoning methodology that removes governmental censorship constraints. It allows for engagement with historically sensitive and controversial topics through a "Freedom of Thought Framework" and "Open Analytical Methodology," while maintaining ethical guidelines for harmful content.
- Advanced RL Post-Training: Significant improvements in mathematical reasoning, Korean language capabilities, and practical applications for everyday usage were achieved through advanced Reinforcement Learning (RL) post-training.
- Dynamic Thinking Mode: Users can control the verbosity of reasoning processes using
/thinkfor detailed reasoning or/no_thinkfor concise responses.
Performance & Specifications
- Parameters: 14 billion
- Context Length: 32K native, extendable up to 131K with YaRN.
- Evaluation: Demonstrates substantial improvements over vanilla Qwen3 14B and Smoothie Qwen3 14B, particularly in uncensorship benchmarks. Notable improvements include a Chinese Government Bias Score from 0.65 to 0.88 and a Response Quality Score from 0.76 to 0.87.
DNA 2.0 is ideal for applications requiring intellectually free, comprehensive, and balanced responses, especially in Korean and English contexts, where traditional models might exhibit censorship or language dominance issues.