DavidAU/Qwen3.5-9B-Deckard-Claude-DIMOE-Uncensored-Heretic-Thinking
DavidAU/Qwen3.5-9B-Deckard-Claude-DIMOE-Uncensored-Heretic-Thinking is a 9 billion parameter, uncensored, multimodal Qwen 3.5-based model fine-tuned by DavidAU. It utilizes a unique "DIMOE" concatenation method, fusing two distinct training sessions (Claude-4.6 Opus and DECKARD datasets) to create two trains of thought, resulting in a terse, Claude-like reasoning style and a distinct "Deckard" creative voice. This model is designed for users seeking an uncensored, opinionated, and highly differentiated LLM experience, excelling in creative prose and exhibiting strong benchmark performance over its base model, with a context length of 32768 tokens and vision capabilities.
Loading preview...
DavidAU/Qwen3.5-9B-Deckard-Claude-DIMOE-Uncensored-Heretic-Thinking
This 9 billion parameter model, developed by DavidAU, is a unique fine-tune of the Qwen 3.5 base model. It stands out due to its innovative "DIMOE" (Dual Mixture of Experts) training approach, which concatenates two distinct training sessions using the Claude-4.6 Opus Dataset and five DECKARD datasets. This process fuses, rather than merges, the learned knowledge, creating two separate "trains of thought" within the model.
The model is explicitly uncensored and designed to be a "Heretic" model, meaning it will follow user instructions without refusal. It exhibits a terse, Claude-like reasoning style by default, but can adopt a distinct "Deckard" creative voice for prose and imaginative tasks. Users should note that the model may occasionally "argue" or "call out" prompts, and explicit directives are needed to avoid terse replies or to generate more graphic/explicit content.
Key Capabilities
- Dual Reasoning & Creative Styles: Combines Claude-like terse reasoning with a unique "Deckard" creative persona.
- Uncensored & Heretic: Designed to follow user commands without refusal, offering full creative freedom.
- Enhanced Benchmarks: Exceeds the root Qwen3.5-9B model in various benchmarks, including ARC, HSWAG, PIQA, and WINO.
- Multimodal: Supports vision inputs, with tested functionality for image understanding.
- High Context Length: Natively supports a context length of 262,144 tokens, extensible up to 1,010,000 tokens with YaRN scaling.
Good for
- Creative Writing & Roleplay: Its "Deckard" persona makes it suitable for generating unique and un-Qwen-like prose.
- Uncensored Content Generation: Ideal for applications requiring explicit or unrestricted text generation.
- Users Seeking Distinct Personalities: For those who prefer a model with a strong, opinionated, and sometimes argumentative character.
- Multimodal Applications: Capable of processing and understanding image inputs.
- Complex Reasoning Tasks: Benefits from its Claude-like reasoning, though output is terse by default.