Impish_Magic_24B: A Specialized 24B Model for Roleplay and Assistant Tasks
Impish_Magic_24B, developed by SicariusSicariiStuff, is a 24 billion parameter model built upon the Mistralai/Magistral-Small-2506 architecture. It underwent a "full" fine-tuning process using over 100 million unique tokens, focusing on modifying vocabulary usage while preserving intelligence. This model is notable for its Apache 2.0 License and its ability to run efficiently on consumer-grade GPUs with 16GB VRAM.
Key Capabilities
- Roleplay Specialization: Features unique and highly specialized data for fighting roleplay, as well as distinct tsundere/yandere interactions. It is highlighted as the first roleplay finetune of Magistral.
- Adventure Generation: Includes new and high-quality adventure data, particularly for settings like Morrowind and Kenshi, with strong recommendations for using specific adventure card syntax.
- Assistant Functionality: Excels as a general assistant, capable of in-depth poem analysis, grammar correction (with detailed explanations), and emulating arbitrary author writing styles. It incorporates extended training data from a previous grammar-focused project.
- Censorship & Tone: Exhibits low to medium censorship (rated 6.5/10) and a slightly reduced positivity bias compared to its base model, influenced by its fighting roleplay data.
- Efficient Performance: Designed to be runnable on mid-tier gaming GPUs (e.g., 4090m) and even high-end mobile devices, offering a balance between intelligence and accessibility.
Good For
- Users seeking a model optimized for detailed and dynamic roleplay scenarios, especially those involving combat or specific character archetypes.
- Developers and enthusiasts looking for a model capable of generating rich, immersive adventure narratives with item tracking and implicit logic.
- Applications requiring a versatile assistant for tasks like advanced grammar checking, creative writing assistance, and complex text analysis.
- Individuals with 16GB VRAM GPUs who need a powerful yet locally runnable model for both creative and analytical tasks.