kawaimasa/Wanabi-Gemma4-31B
kawaimasa/Wanabi-Gemma4-31B is a 31 billion parameter model based on Google's Gemma 4 31B-IT, fine-tuned for creative writing tasks in Japanese. It is optimized to understand and respond to structured prompt formats from the Project Wannabe desktop application, while also supporting general chat UIs. This model integrates Japanese Chain-of-Thought (CoT) data to enhance natural Japanese expression and writing style in creative tasks, maintaining the base model's general dialogue and reasoning capabilities.
Loading preview...
Wanabi-Gemma4-31B: Optimized for Japanese Creative Writing
Wanabi-Gemma4-31B is a 31 billion parameter model, fine-tuned from Google's Gemma 4 31B-IT. Its primary distinction lies in its optimization for Japanese creative writing tasks, particularly for the dedicated desktop application Project Wannabe. Unlike previous Wanabi models, this version leverages an Instruct model base, allowing it to retain strong general dialogue and reasoning capabilities while excelling in creative generation.
Key Capabilities
- Project Wannabe Workflow: Seamlessly integrates with Project Wannabe for structured novel generation (GEN), continuation (CONT), and idea generation (IDEA) based on unique structured prompt formats.
- Enhanced Japanese Expression: Incorporates Japanese Chain-of-Thought (CoT) data to improve the naturalness of Japanese vocabulary, style, and overall creative expression, mitigating the "translation-like" stiffness often seen when models reason in English.
- Versatile UI Support: Can be used with general chat UIs like OpenWebUI and SillyTavern for natural Japanese dialogue, role-playing, and free-form creative writing, a significant improvement over previous Project Wannabe-exclusive models.
What Makes It Different?
This model represents a shift in approach from prior Wanabi series. Instead of building writing capabilities from a Base model, it fine-tunes an Instruct model, preserving its broad knowledge while gently steering its output towards creative tasks. The focus on Japanese CoT data specifically aims to make the model "think" in Japanese for creative tasks, leading to more authentic and nuanced output.
Good For
- Novelists and Writers: Especially those using the Project Wannabe application for structured story creation.
- Role-playing Enthusiasts: Provides natural and expressive Japanese for character interactions.
- General Japanese Chat: Offers robust dialogue capabilities in Japanese.
Note: While it supports a 32K context length, training was primarily up to 24K. For coding or structured formats like JSON, the base Gemma 4 31B-IT might be preferred due to potential formatting issues from the Japanese CoT data translation process.