CtrlCreeper/Qwen3.8-27B-Cold-Fusion-Heretic-Fable-Heretic
CtrlCreeper/Qwen3.8-27B-Cold-Fusion-Heretic-Fable-Heretic is a 27 billion parameter Qwen3.8-based model, created by CtrlCreeper, that merges two 'abliterated' Qwen3.8 derivatives. It combines the reasoning and concise behavior of Cold-Fusion GAIN with the creative and conversational characteristics of Fable-Distill. This model is designed as a general-purpose Qwen3.8 derivative, offering a balance between strong general reasoning and creative writing capabilities, while retaining the native vision-language tower and MTP speculative-decoding head of the Qwen3.8 architecture.
Loading preview...
Overview
CtrlCreeper/Qwen3.8-27B-Cold-Fusion-Heretic-Fable-Heretic is a 27 billion parameter model built upon the Qwen3.8 architecture, created by merging two specialized Qwen3.8 derivatives: gorbatjovy/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-heretic and armand0e/Qwen3.8-27B-Fable-Distill-Heretic-ara. This merge utilizes a NuSLERP method with a dominant weighting towards the Cold-Fusion component, aiming to blend its strong general reasoning and concise task-oriented responses with Fable-Distill's creative writing, natural dialogue, and narrative style.
Key Characteristics
- Hybrid Architecture: Retains the Qwen3.8 hybrid architecture, including 64 language-model layers, hybrid linear + full attention, a native vision-language tower, and an MTP speculative-decoding head.
- Abliterated Sources: Both source models have undergone 'abliteration' (refusal-direction removal), meaning this merged model may comply with prompts that official aligned Qwen3.8 models would refuse. Users are advised to implement appropriate moderation.
- Balanced Capabilities: The merge is weighted to prioritize reasoning and general capability while still incorporating creative and conversational elements.
- Tensor-by-Tensor Merge: The merge was performed directly tensor-by-tensor across all 1199 matching tensors to accurately preserve the Qwen3.8's hybrid attention layout.
Intended Use Cases
This model is designed as a versatile, general-purpose Qwen3.8 derivative suitable for applications requiring:
- Reasoning and General Tasks: Benefits from the Cold-Fusion component for concise and task-oriented responses.
- Creative Writing and Dialogue: Leverages the Fable-Distill component for generating natural dialogue, character voices, and narrative content.
- Vision-Language Tasks: Inherits the native vision-language tower from the Qwen3.8 base.
Available in various quantizations including BF16, MLX MXFP8/MXFP4, and GGUF (Q4_K_M, Q6_K, Q8_0).