nightmedia/Qwen3.6-27B-Miraculix
nightmedia/Qwen3.6-27B-Miraculix is a 27 billion parameter language model, a NuSLERP merge of several Qwen3.6-27B variants, including nbeerbower/Wichtel-Qwen3.6-27B and EpistemeAI/Reasoning-Medical-27B. This model achieves an ARC score of 0.743 on mxfp8 quantization, an unprecedented score for its parameter size, and features a unique 'think' tag for self-correction and mental simulation. It is optimized for advanced reasoning and contextual understanding, particularly in complex, multi-step tasks, and integrates medical reasoning capabilities.
Loading preview...
Model Overview
nightmedia/Qwen3.6-27B-Miraculix is a 27 billion parameter language model, created through a NuSLERP merge of several specialized Qwen3.6-27B models, including nbeerbower/Wichtel-Qwen3.6-27B and EpistemeAI/Reasoning-Medical-27B. This unique combination allows it to achieve exceptional reasoning capabilities, evidenced by an ARC score of 0.743 on mxfp8 quantization, which is noted as unprecedented for its parameter footprint.
Key Capabilities
- Advanced Reasoning: Excels in complex logical tasks, with a high ARC benchmark score.
- Self-Correction Mechanism: Utilizes a novel
<|think_on|>and<|think_off|>tag system for internal mental simulation and self-refinement during inference, enhancing output quality. - Medical Reasoning Integration: Incorporates specialized knowledge from a medical reasoning model, allowing for nuanced advice and understanding in health-related contexts.
- Efficient Quantization: Demonstrates strong performance across various quantization levels (mxfp8, q8-hi, q6-hi, mxfp4) with optimized memory usage.
What Makes it Different
This model's primary differentiator is its "GrandWichtel architecture," which enables it to surpass perceived architectural limits for Qwen3.6-27B models in reasoning tasks. The integrated 'think' tag, while potentially increasing inference time due to internal mental simulation, allows for a deeper, more robust problem-solving approach. Its blend of general reasoning with specialized medical knowledge, combined with its high benchmark performance for its size, positions it as a powerful tool for applications requiring both analytical depth and contextual understanding.