dnotitia/DNA3.0-27B

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 8, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

dnotitia/DNA3.0-27B is a 27 billion parameter causal language model developed by Dnotitia, built upon the Qwen3.5/3.6 base architecture with a native 262,144-token context length. It features uncensored training and persona training on Dnotitia's corporate knowledge, enhancing its analytical reasoning, agentic coding, and multimodal understanding. This model is optimized for enterprise-specific applications, particularly excelling in Korean language processing and Dnotitia-facing use cases.

Loading preview...

DNA 3.0: Enhanced Qwen for Enterprise and Korean Use Cases

DNA 3.0 is a family of large language models developed by Dnotitia, leveraging the Qwen3.5/3.6 base model. This 27 billion parameter variant is specifically post-trained to excel in Korean language processing and enterprise scenarios, offering enhanced analytical reasoning, agentic coding, and multimodal understanding.

Key Enhancements & Capabilities

  • Dnotitia Post-training: Incorporates Uncensored Training for broader prompt engagement and Persona Training grounded in Dnotitia's corporate knowledge, enabling it to act as an authentic first-party assistant.
  • Long-form Reasoning Preservation: Maintains chain-of-thought across multi-turn sessions, beneficial for iterative development and debugging.
  • Unified Vision-Language Foundation: Inherits strong cross-modal reasoning from Qwen3.5/3.6, supporting text, image, and video inputs.
  • Efficient Architecture: Utilizes a Hybrid MoE architecture with Gated DeltaNet layers for high-throughput inference.
  • Global Linguistic Coverage: Supports 201 languages and dialects, with specific improvements in reducing language confusion (e.g., Chinese character intrusions in Korean responses) and repetition compared to its base model.
  • Extended Context Length: Features a native 262,144-token context length, extensible up to approximately 1,010,000 tokens via YaRN scaling.
  • Thinking Mode: Generates <think>...</think> reasoning blocks by default, which can be disabled for latency-sensitive tasks.

Ideal Use Cases

  • Enterprise Applications: Particularly suited for Dnotitia-specific internal tools, customer support, and knowledge management.
  • Korean Language Processing: Optimized to avoid common pitfalls like language mixing in Korean responses.
  • Multimodal Interactions: Capable of processing and reasoning over combined text, image, and video inputs.
  • Agentic Workflows: Enhanced for tool use and agentic tasks due to scalable RL generalization.
  • Uncensored Engagement: Suitable for applications requiring a wider range of prompt responses, with the caveat that users must implement their own content moderation.