hiwaifu-research/WaifuGemma4-26b-a4b-v1
WaifuGemma4-26b-a4b-v1 by HiWaifu Research is a 26 billion parameter Mixture of Experts (MoE) language model, with 3.8 billion active parameters, built on the Gemma 4 architecture. It is post-trained using GRPO on 1.2 million real user preference votes from the HiWaifu role-play arena, making it highly optimized for engaging and creative role-play conversations. This model demonstrates performance comparable to GLM-5.1 in blind head-to-head user evaluations, excelling in generating expressive and proactive replies across multiple languages with a 32K token context length.
Loading preview...
WaifuGemma4-26b-a4b-v1: Role-play Optimized Gemma 4 MoE
WaifuGemma4-26b-a4b-v1 is a 26 billion parameter (3.8B active) Mixture of Experts model from HiWaifu Research, based on the Gemma 4 architecture. Its key differentiator is its unique post-training process: 200 GRPO steps using a reward model derived from 1.2 million real user votes collected in the HiWaifu role-play arena. This method ensures the model's outputs are directly aligned with actual user preferences in dynamic, multi-turn role-play scenarios.
Key Capabilities & Performance
- Human Preference Alignment: Achieves a 54.7% win rate against a 13-model field in blind arena battles and is rated dead even with GLM-5.1 (49.6% win rate head-to-head).
- Role-play Specialization: Excels in generating creative, emotionally deep, and proactive replies, with win rates increasing in longer conversations (56.1% from turn 11 onwards).
- Multilingual Support: Evaluated across Spanish, Russian, English, Portuguese, Indonesian, Arabic, and Thai, demonstrating strong performance in diverse languages.
- General Ability Preservation: Maintains general benchmark performance, with only a -0.3 point average change across 10 benchmarks compared to its untuned base, indicating specialized fine-tuning without significant degradation of core capabilities.
Ideal Use Cases
- Creative Role-play Applications: Best suited for interactive storytelling, character-driven chatbots, and virtual companions where engaging and expressive dialogue is paramount.
- Multilingual Conversational Agents: Effective in scenarios requiring nuanced, preference-aligned responses in various languages.
- Applications requiring Human-Centric Outputs: Where user satisfaction and natural, preferred conversational flow are critical, especially in long-form interactions.