trinityomni/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU
The trinityomni/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU is a 27 billion parameter Qwen3.8-based language model developed by trinityomni, fine-tuned for enhanced reasoning, reduced token usage for thinking, and uncensored output. It demonstrates significant performance improvements over base Qwen 3.8 models, particularly in ARC-C benchmarks, and maintains strong performance even in 4-bit quantization. This model is optimized for detailed analytical tasks and creative generation, offering stable performance across various quantization levels.
Loading preview...
trinityomni/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU
This model is a 27 billion parameter variant based on the Qwen 3.8 architecture, developed by trinityomni. It has undergone extensive multi-stage fine-tuning, including "heretic'ing" (uncensoring) and subsequent "healing" training to restore and enhance core metrics. The model is designed to offer superior performance compared to other Qwen 27B models and even some 35B models, with a focus on reasoning, reduced 'thinking' token count, and uncensored output.
Key Capabilities
- Enhanced Reasoning: Demonstrates a significant rise in core metrics, with ARC-C scores 141 points above the base Qwen 3.8 27B benchmark. This includes specialized reasoning adjustments in later development branches.
- Efficient Thinking: Achieves a strong reduction in 'overthinking' or 'thinking tokens' (1/2 to 1/10 of normal Qwen size) while maintaining a high level of detail. It features auto-variable thinking sizes based on prompt and use case.
- Uncensored Output: The model has been "heretic'ed" for uncensored responses, balanced with a low KLD (Kullback-Leibler Divergence) to minimize damage to core performance.
- Quantization Stability: Maintains strong performance in 4-bit quantization, with benchmarks sitting at or close to 99% of 8-bit performance, ensuring rock-solid stability.
- Creative and Detailed Generation: Exhibits strong creative performance upgrades and the ability to produce detailed, analytical, and even emotionally resonant outputs, as shown in example generations.
Good For
- Advanced Reasoning Tasks: Ideal for applications requiring deep analytical thought and complex problem-solving, where traditional models might overthink or provide less concise reasoning.
- Creative Writing and Roleplay: Its uncensored nature and enhanced creative capabilities make it suitable for generating diverse and imaginative content, including fiction and character-driven narratives.
- Resource-Efficient Deployment: Given its stable performance in 4-bit quantization, it's well-suited for environments where computational resources are a concern, allowing for high performance with reduced memory footprint.
- Applications Requiring Direct and Unfiltered Responses: For use cases where strict content filtering is not desired, and direct, unconstrained output is preferred.