YFC-112358/Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 18, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

YFC-112358/Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4 is a 27 billion parameter Qwen3.8-based model created by YFC-112358, developed through a three-stage merging process. This model integrates multiple Qwen3.6-based components, including DELLA, Qwopus3.6-27B-Fusion, and Cold-Fusion-GAIN-V1.1, to create a "general intelligent composite." It is specifically designed to combine and transfer task vectors across different Qwen generations, aiming for enhanced performance by leveraging diverse foundational models.

Loading preview...

Model Overview

This model, YFC-112358/Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4, is a 27 billion parameter language model built upon the Qwen3.8 architecture. It is the result of a sophisticated three-stage merging process, designed to synthesize capabilities from several Qwen3.6-based models and then "re-anchor" them to a Qwen3.8 base.

Key Merging Stages:

  • Stage 1: General Intelligent Composite (G): Three distinct Qwen3.6-based models (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 merged using a della_linear method to form a composite model 'G'.
  • Stage 2: Qwopus Integration: The 'G' composite is then merged with KyleHessling1/Qwopus3.6-27B-Fusion-BF16 using task_arithmetic, with Qwopus playing a dominant role. This stage is noted to be a pure linear superposition due to specific density parameters.
  • Stage 3: Cross-Generational Re-anchoring: The output from Stage 2 is then re-anchored to DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1. This involves linearly transferring task vectors relative to Qwen3.6 onto the Qwen3.8-based Cold-Fusion model. This "cross-generational linear transfer" is highlighted as the model's most significant uncertainty, as its validity depends on the alignment of the two generations' bases.

Unique Characteristics:

  • Complex Multi-Stage Merging: Utilizes advanced merging techniques like DELLA and task arithmetic across multiple foundational models.
  • Cross-Generational Task Vector Transfer: Aims to port learned capabilities from Qwen3.6 derivatives to a Qwen3.8 base, a method with inherent experimental aspects.
  • Configurable Adapter: A complementary LoRA adapter (YFC-112358/Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4-LoRA) is available, allowing for flexible integration with other Qwen3.8 models.

Usage Considerations:

  • The model's performance relies on the assumption that Qwen3.6 and Qwen3.8 share a sufficiently aligned base for task vector transfer. Empirical cold_amp readings suggest this might be within acceptable experimental ranges.
  • Sampling recommendations include temperature=0.7, top_p=0.8, top_k=20.

Finetunes
1 models