USS-Inferprise/Dark-Cydonian-Wind-24B
Dark Cydonian Wind is a 24 billion parameter TIES merge model based on three Mistral-Small derivatives, optimized for creative writing with a nuanced 'spiciness' and reduced 'slop'. It supports large context windows of 48K-64K tokens and is designed to fit on 24GB VRAM cards. This model excels at generating engaging narrative content, particularly for creating complex characters and scenarios.
Loading preview...
Dark Cydonian Wind: A Nuanced 24B Merge Model
Dark Cydonian Wind is a 24 billion parameter model created through a TIES merge of three Mistral-Small derivatives: Cydonia-24B-v4.3 (45%), The-Omega-Darker-The-Final-Directive-24B (30%), and Redemption Wind (25%). This unique combination aims to balance narrative quality with a distinct 'spiciness' while maintaining stability and reducing undesirable 'slop' often found in parent models.
Key Capabilities
- Large Context Support: Capable of handling extensive context windows ranging from 48K to 64K tokens.
- VRAM Efficiency: Designed to operate effectively on GPUs with 24GB VRAM, and even 16GB VRAM at Q4_K_M quantization with a reduced context.
- Balanced 'Spiciness': Offers an 'edge' suitable for creative writing, enabling the generation of compelling, morally ambiguous characters and scenarios without crossing into extreme or inappropriate content.
- Improved Coherence: The TIES merge process has significantly reduced 'slop' from its constituent models, leading to more coherent and less erratic outputs.
Good For
- Creative Writing: Ideal for generating nuanced narratives, character monologues, and complex plot developments, especially when a touch of 'villainy' or moral ambiguity is desired.
- Roleplay Scenarios: Can create engaging and dynamic interactions for role-playing applications where characters with depth and a slight edge are beneficial.
[!WARNING]
This model incorporates components with weakened or removed refusal vectors. Users should exercise caution as it may produce content unsuitable for minors or vulnerable individuals.