OpceanAI/Yuuki-RxG
OpceanAI/Yuuki-RxG is an 8 billion parameter language model fine-tuned from DeepSeek-R1-Distill-Qwen-8B, featuring a 32,768 token context length. It is specialized for advanced reasoning and competition-level mathematics, significantly outperforming its base model and achieving a 96.6% score on TruthfulQA. This model excels in complex problem-solving and verifiable factual honesty, making it suitable for applications requiring high cognitive rigor.
Loading preview...
YuuKi RxG: Advanced Reasoning and Factual Honesty
YuuKi RxG is OpceanAI's flagship 8 billion parameter language model, built upon the DeepSeek-R1-Distill-Qwen-8B base. It represents the first release in the RxG family, specifically engineered for advanced reasoning, mathematical rigor, and verifiable factual honesty. The model significantly surpasses its base model across various benchmarks, including AIME 2024, AIME 2025, HMMT February 2025, GPQA Diamond, and LiveCodeBench.
Key Capabilities & Differentiators
- Exceptional Factual Honesty: Achieves an unprecedented 96.6% on TruthfulQA, the highest known for any open-weight model, emerging naturally from its training rather than explicit honesty instruction.
- Competition-Level Mathematics: Demonstrates strong performance in advanced mathematics, competitive with models an order of magnitude larger, such as o3-mini and Gemini-2.5-Flash-Thinking.
- Advanced Reasoning: Inherits and preserves the native
<think>block protocol from its DeepSeek-R1 base, allowing it to explicitly reason before responding, which is a core property of the model. - Consistent Identity: Maintains the warm, bilingual (English, Spanish) identity characteristic of the YuuKi model family, trained directly into its weights.
- Robust Architecture: Fine-tuned using Supervised SFT + LoRA, supporting a 32,768 token context length, though fine-tuning was primarily conducted at 4,096 tokens.
Ideal Use Cases
- Complex Problem Solving: Suited for tasks requiring deep logical inference and structured reasoning.
- Mathematical Applications: Excellent for competition mathematics, formal proofs, and quantitative analysis.
- Fact-Checking & Information Retrieval: Its high TruthfulQA score makes it valuable for applications where factual accuracy is paramount.
- Educational Tools: Can be used for generating explanations and step-by-step reasoning processes in academic contexts.
While strong in reasoning and honesty, the model has identified limitations in graduate-level science reasoning (GPQA Diamond) and code generation compared to specialized models. Performance beyond 4,096 tokens context has not been formally evaluated.