Naphula/StormSeeker-24B-v1

StormSeeker-24B-v1 Overview

StormSeeker-24B-v1 is a 24 billion parameter merged language model developed by Naphula, utilizing a custom flux merge method. This model integrates six distinct base models, with Loki, PaintedFantasy, and Hearthfire having a notable influence on its characteristics. It is specifically designed for generating narratives and roleplay content, including potentially violent and graphic erotic material, and operates with a largely uncensored output.

Key Characteristics & Performance

• Uncensored Output: The model is noted for its uncensored nature, responding to some harmful prompts without refusals. A light jailbreak can effectively bypass most censorship.
• Optimized Prompting: For optimal performance, users are advised to use Mistral non-Tekken prompting, which yields a significantly higher score (8867) compared to Tekken prompting (6933).
• Merge Method: The model was created using a flux merge method (version 5, Y6 config), which involved 1005 iterations to achieve maximum BF16 fidelity.
• Quantization Benefits: Merges made with flux method are suggested to benefit greatly from smaller block sizes like IQ4_NL quantizations, potentially performing on par with or better than Q6_K, though this claim is not yet empirically verified.

Use Cases & Considerations

• Narrative and Roleplay: Ideal for applications requiring creative writing, detailed narratives, and roleplay scenarios, especially those that may involve graphic or uncensored content.
• Content Warning: Users should be aware of its capacity to produce violent and graphic erotic content and adjust system prompts accordingly.
• Ablation: An MPOA-Adapter is available for ablations to modify the model's behavior.

Available Quantizations

Due to constraints, the model is primarily uploaded in IQ4_NL, Q6_K, and Q8_K_XL GGUF formats. Recommendations for other quantizations (e.g., IQ4_XS, EXL3, MLX) are provided via links to other community members.