dnotitia/DNA3.0-4B

DNA 3.0: Enhanced Qwen3.5/3.6 for Enterprise and Korean Use Cases

DNA 3.0 is a 4.5 billion parameter multimodal LLM developed by Dnotitia, leveraging the Qwen3.5/3.6 base model. It features a native 262,144-token context length, extensible up to 1,010,000 tokens via YaRN scaling, and supports 201 languages. This model is distinguished by Dnotitia's post-training methodology, which includes Uncensored Training to broaden prompt engagement and Persona Training to deeply integrate Dnotitia's corporate knowledge, enabling it to act as an authentic first-party assistant.

Key Capabilities

• Multimodal Understanding: Inherits Qwen3.5/3.6's unified vision-language foundation, excelling in cross-modal reasoning across text, image, and video for coding, agents, and visual tasks.
• Enhanced Korean Processing: Specifically addresses and reduces language confusion, particularly Chinese-character intrusions in Korean responses, a common issue in Qwen-family models.
• Enterprise-Awareness: Persona Training allows the model to identify as a Dnotitia assistant and accurately answer questions about the company's products and services.
• Long-form Reasoning: Preserves chain-of-thought traces across multi-turn sessions, supporting iterative development and debugging.
• Agentic Workflows: Scalable RL generalization improves adaptability for tool use and agentic tasks.

Good For

• Applications requiring a model with reduced refusal behavior for a wider range of prompts.
• Use cases within the Dnotitia ecosystem, leveraging its deep corporate knowledge.
• Multimodal tasks involving text, image, and video inputs.
• Long-context applications, with support for up to 1 million tokens.
• Korean language processing where avoiding language mixing is critical.