YFC-112358/Qwen3.6-27B-Della-Deckard-Fable-Qwopus-v3
YFC-112358/Qwen3.6-27B-Della-Deckard-Fable-Qwopus-v3 is a 27 billion parameter language model based on the Qwen3.6 architecture, developed by YFC-112358. This model is a sophisticated merge of multiple Qwen3.6 variants, utilizing a two-stage merging process that combines DELLA and task arithmetic methods to create a "general intelligence composite." It is designed to integrate diverse model strengths, particularly focusing on preserving unique features from its constituent models, and is suitable for general-purpose language generation tasks.
Loading preview...
Model Overview
YFC-112358/Qwen3.6-27B-Della-Deckard-Fable-Qwopus-v3 is a 27 billion parameter model built upon the Qwen3.6 architecture. It is the result of a two-stage merging process designed to synthesize the strengths of several distinct Qwen3.6-based models into a "general intelligence composite." The merging strategy ensures that the incremental improvements from each contributing model are linearly combined, maintaining their unique characteristics.
Merging Methodology
The model employs a sophisticated two-stage merging approach:
- Stage 1 (DELLA): Three different Qwen3.6-27B variants (DavidAU/Qwen3.6-27B-V1.1-FF711-Darker-Hero-GAIN-H2.0, YFC-112358/Qwen3.6-27B-Della-Deckard-Isometry-Geodesic-v2, and nightmedia/Qwen3.6-27B-Seven) are fused using the DELLA method to create an intermediate "general intelligence composite" (G).
- Stage 2 (Task Arithmetic): The incremental changes from KyleHessling1/Qwopus3.6-27B-Fusion-BF16 relative to the base Qwen3.6-27B are linearly added to the composite G. This stage specifically uses task arithmetic, diverging from DELLA to avoid pruning small, yet significant, increments.
Key Characteristics
- Preservation of Features: The merging process is carefully designed to avoid pruning or discarding unique features from the constituent models, especially by opting for linear superposition in the second stage and disabling TIES sign election where it might project away orthogonal characteristics.
- Efficiency: The two-stage calculation is performed in a single block without saving the intermediate
stage1-Gmodel to disk, saving significant storage and transfer overhead. - Robustness: Epsilon values for density are automatically narrowed to legal limits, ensuring mergekit compatibility while preserving intended density.
Usage
This model can be loaded using the Hugging Face transformers library. Recommended sampling parameters include temperature=0.7, top_p=0.8, and top_k=20 for optimal generation quality.