Klear-Reasoner - The New Reasoning Model Introduced by Racer

What is Klear-Reasoner?

Klear-Reasoner is a high-performance inference model introduced by Racer and developed based on Qwen3-8B-Base. The model is trained through long thought chain supervised fine-tuning and reinforcement learning, and excels in mathematical and coded reasoning.Klear-Reasoner's core innovation is the GPPO algorithm, which significantly improves the model's exploration ability and convergence speed of negative samples based on preserving the cropped gradient information. In benchmarks such as AIME and LiveCodeBench, Klear-Reasoner demonstrates excellent performance, reaching the top level of 8B models. With its ability to solve complex mathematical problems and generate high-quality code snippets, Klear-Reasoner is widely used in education, software development, and fintech, providing a valuable reference and reproducible path for the development of inference models.

Features of Klear-Reasoner

• mathematical reasoning: Models excel at solving complex math problems, providing students with clear solutions and steps to help users better understand and master math.
• Code Generation and Reasoning: Generates high-quality code snippets that assist developers in quickly implementing functional modules.
• long thought chain reasoning: Supervised fine-tuning and reinforcement learning based on long chains of thought can improve the performance of models in multi-step reasoning and support the handling of complex reasoning tasks.
• Data quality optimization: Prioritize high-quality data sources during the training process, while retaining some error samples to enhance the model's exploratory capability.

Klear-Reasoner's Core Benefits

• Efficient training methods: Combining long thought chain supervised fine-tuning and reinforcement learning to fully utilize the advantages of both, the model excels on complex reasoning tasks and lays the foundation for high-performance reasoning.
• Innovative GPPO algorithm: Decoupling clip and gradient backpropagation through stop gradient operation, retaining all token gradient information, improving model exploration capability and negative sample convergence speed, and significantly optimizing training efficiency.
• Powerful reasoningThe model excels in mathematical and code reasoning, solves difficult mathematical competitions and generates high-quality code snippets, which is applicable to education, software development and other fields, and has a promising application prospect.
• Balance between data quality and exploratory capacity: The model prioritizes high-quality data sources while retaining some error samples to enhance exploration.
• Open Source and Reproducibility: The training details and full process of Klear-Reasoner are openly available, and open source resources and detailed documentation are provided to promote academic exchanges and technological advances.

What is Klear-Reasoner's official website?

• GitHub repository:: https://github.com/suu990901/KlearReasoner/
• HuggingFace Model Library:: https://huggingface.co/Suu/Klear-Reasoner-8B
• arXiv Technical Paper:: https://arxiv.org/pdf/2508.07629

Who is Klear-Reasoner for?

• schoolchildren: Students will be able to solve math puzzles, obtain detailed step-by-step solutions, and gain a better understanding and mastery of mathematics
• software developer: Software developers generate high-quality code snippets, quickly implement functional modules, and improve development efficiency and code quality.
• Fintech practitioners: FinTech practitioners analyze financial data for risk assessment and prediction to help make more accurate decisions.
• (scientific) researcher: Researchers deal with complex data analysis to obtain logical reasoning and improve research efficiency.
• Intelligent Customer Service Team: Intelligent customer service teams answer complex user questions quickly and accurately, improving user experience and problem-solving efficiency.