dnotitia/DNA3.0-35B-A3B
dnotitia/DNA3.0-35B-A3B is a 35.1 billion parameter Mixture-of-Experts model developed by Dnotitia, built upon the Qwen3.5/3.6 base. It features uncensored training and persona training for Dnotitia's corporate knowledge, excelling in analytical reasoning, agentic coding, and multimodal understanding. This model is optimized for enterprise-aware conversational capabilities, particularly for Korean language use cases, while maintaining a 32K token context length.
Loading preview...
DNA 3.0: Enhanced Qwen3.5/3.6 for Enterprise and Korean
dnotitia/DNA3.0-35B-A3B is a 35.1 billion parameter Mixture-of-Experts (MoE) model, extending the Qwen3.5/3.6 base with specialized post-training by Dnotitia. It features an efficient hybrid MoE architecture with 256 experts (8 routed + 1 shared) and a native context length of 262,144 tokens, extensible up to 1,010,000 tokens.
Key Enhancements & Capabilities
- Uncensored Training: Provides a broader range of responses without unnecessary refusals, preserving reasoning quality.
- Persona Training: Grounded in Dnotitia's corporate knowledge, enabling it to act as an authentic first-party assistant.
- Long-form Reasoning Preservation: Retains chain-of-thought across multi-turn sessions for complex workflows.
- Unified Vision-Language Foundation: Delivers strong cross-modal reasoning across text, image, and video, outperforming prior Qwen3-VL models.
- Global Linguistic Coverage: Native support for 201 languages and dialects, with specific improvements for Korean by reducing language confusion and repetition.
- Thinking Mode by Default: Generates
<think>...</think>reasoning blocks, which can be disabled for latency-sensitive tasks.
Differentiators from Qwen3.6-35B-A3B
DNA3.0-35B-A3B shows significant improvements over its base model in:
- Persona Identification: Reliably identifies as a Dnotitia assistant.
- Uncensorship: Engages with topics the Chinese-origin base model typically refuses.
- Language Confusion Reduction: Minimizes unintended language mixing, especially Chinese characters in Korean responses.
- Repetition Reduction: Avoids infinite-loop repetition during long-form generation.
Use Cases
This model is ideal for enterprise applications requiring robust analytical reasoning, agentic coding, and multimodal understanding, particularly in Korean-speaking markets or for Dnotitia-specific internal tools. Its uncensored nature allows for wider prompt engagement, though users should implement appropriate content moderation.