DragonBophades/Elster-Vernunft-Qwen3.6-27B
DragonBophades/Elster-Vernunft-Qwen3.6-27B is a 27 billion parameter Qwen3.6-based model, created by DragonBophades, that integrates the Vernunft reasoning LoRA at full scale. This model significantly reduces the length of internal reasoning by 8.4 times and eliminates first-person self-talk, while maintaining performance on tasks like ARC-Challenge and coding. It is optimized for efficient and structured reasoning, making it suitable for applications requiring concise logical thought processes.
Loading preview...
Elster-Vernunft-Qwen3.6-27B: Efficient Reasoning Model
This model, developed by DragonBophades, is a 27 billion parameter Qwen3.6 variant that merges the Vernunft-Qwen3.6-27B-LoRA at full 1.0 scale. It is designed to drastically reduce the verbosity of the model's internal reasoning processes without compromising accuracy on general reasoning tasks.
Key Capabilities & Differentiators
- 8.4x Shorter Reasoning: Achieves an 8.4 times reduction in the median length of internal reasoning compared to its base model, Elster.
- Eliminated R1 Self-Talk: Completely removes first-person self-talk (R1 register) from its reasoning output, resulting in 100% structured reasoning.
- Performance Preservation: Maintains performance on ARC-Challenge and Hemlock coding tasks, showing no measurable degradation despite the reasoning compression.
- Full-Scale LoRA Integration: Unlike previous attempts, the Vernunft adapter is integrated at full scale due to its rewritten training data, which prevents performance cliffs.
Important Considerations & Limitations
- Not for Agentic Tool-Calling: This model is explicitly not recommended for agentic tool-calling applications, as its accuracy in tool selection and argument parsing is significantly reduced.
- Context-Dependent Reasoning Length: While generally shorter, reasoning length may slightly increase for very short, two-sentence prompts where there is little to compress.
This model is ideal for use cases where concise, structured, and efficient reasoning is paramount, and where agentic tool-use is not a primary requirement.