Yuuki NxG: A 3B Companion Model
Yuuki NxG is a 3-billion parameter language model developed by OpceanAI, fine-tuned from the Qwen2.5-3B architecture. This model is notable for being trained entirely on a Mac Pro (2020) with no cloud compute budget, showcasing that meaningful alignment fine-tuning is achievable on consumer hardware.
Key Capabilities & Features
- Personality Alignment: Fine-tuned for consistent, context-aware conversation, excelling in emotional support and casual Q&A.
- Factual Honesty: Achieves the highest TruthfulQA score (50.87%) among compared 3B-scale models, including its base model, indicating improved factual calibration through fine-tuning.
- Zero-Budget Training: Developed without any cloud compute expenditure, demonstrating an accessible approach to AI development.
- Multilingual: Functional in both English and Spanish, inheriting capabilities from its Qwen2.5 base.
- Open Source: Released under Apache 2.0, allowing commercial use, modification, and distribution.
Performance Highlights
Despite being evaluated 0-shot against competitors using 5–25 shot prompting, Yuuki NxG demonstrates strong performance, particularly in social sciences and humanities. Its MMLU score is 60.65%, with high scores in Marketing (87.18%) and High School Psychology (83.67%). Degradation in HellaSwag is an expected tradeoff for personality-aligned models.
Intended Use Cases
- General-purpose conversational assistance.
- Emotional support and companionship applications.
- Educational Q&A in humanities and social sciences.
- Research into small-scale fine-tuning and personality alignment.
- Local deployment on consumer hardware.