OpceanAI/Yuuki-RxG-nano
OpceanAI's Yuuki-RxG-nano is a 1.5 billion parameter reasoning-specialized language model, fine-tuned from VibeThinker-1.5B, a Qwen2.5-Math architecture. It excels in mathematical and general reasoning tasks, achieving 80.0% on AIME 2024 and 89.6% on TruthfulQA, outperforming larger models in specific benchmarks. Designed for edge deployment, it integrates a consistent identity with strong reasoning capabilities, making it suitable for applications requiring efficient, accurate logical processing.
Loading preview...
YuuKi RxG Nano: Edge Reasoning at 1.5B Scale
YuuKi RxG Nano, developed by OpceanAI, is a 1.5 billion parameter language model specialized in reasoning, built upon the VibeThinker-1.5B base. This model is a compact, edge-deployment entry in OpceanAI's RxG family, designed to deliver high reasoning capability within a small footprint.
Key Capabilities & Performance
- Exceptional Reasoning: Achieves 80.0% on AIME 2024, significantly outperforming other 1.5B models like DeepSeek-R1-Distill-1.5B (28.9%).
- High Truthfulness: Scores 89.6% on TruthfulQA (1-shot), demonstrating strong factual accuracy.
- General Knowledge: Surpasses DeepSeek V3 671B on MMLU-Pro (65.63% vs 64.4%) despite being 447 times smaller.
- Efficient Training: Fine-tuned in approximately 90 minutes on a single GPU for under $15, validating its cost-effectiveness.
- Native Thinking Protocol: Preserves the VibeThinker base's
<think>blocks, reflecting genuine intermediate computation during inference. - Bilingual Support: Responds natively in English or Spanish without explicit instruction.
What Makes It Different?
RxG Nano's unique approach involves inheriting robust reasoning from the VibeThinker base (a distillation of frontier systems like Claude, Gemini, and Kimi) and then installing a consistent "YuuKi" identity via a lightweight LoRA fine-tuning pass. This method ensures reasoning capabilities are preserved while integrating a distinct personality. The model's ability to achieve competitive benchmarks against much larger models, particularly in mathematics and general reasoning, at such a small scale and low training cost, highlights its efficiency and advanced distillation techniques.
Ideal Use Cases
- Edge Deployment: Optimized for environments with limited computational resources.
- Mathematical & Logical Reasoning: Excellent for tasks requiring precise problem-solving and formal reasoning.
- Factual Q&A: Strong performance in truthfulness and general knowledge makes it suitable for accurate information retrieval.
- Interactive Agents: Its consistent identity and explicit reasoning process (via
<think>blocks) can enhance conversational AI applications.