Klear-Reasoner - Kuaishou's open-source inference model

What is Klear-Reasoner?

Klear-Reasoner is an inference model launched by Kuaishou based on Qwen3-8B-Base, focusing on improving mathematical and code reasoning capabilities. The model is trained through long thought chain supervised fine-tuning (long CoT SFT) and reinforcement learning (RL). Its core innovation is the GPPO algorithm, which addresses the limitations of traditional methods in exploration and slow convergence to negative samples by preserving pruned gradient information. It achieves top-tier performance among 8B models in benchmarks such as AIME and LiveCodeBench. The training details and full workflow of Klear-Reasoner are publicly available, providing important reference and replication paths for the development of inference models.

Klear-Reasoner's main functions

• Mathematical reasoningKlear-Reasoner excels at solving complex mathematical problems and can solve highly challenging math competition problems.
• Code generation and reasoningIt can generate high-quality code, achieving accuracy rates of 66.0% and 58.1% respectively in LiveCodeBench V5 and V6 tests.
• Long-chain reasoningKlear-Reasoner can handle complex long thought chain tasks. It improves the model's performance in multi-step reasoning through long thought chain supervised fine-tuning (long CoT SFT) and reinforcement learning (RL) training.
• Data quality optimizationDuring training, Klear-Reasoner prioritizes high-quality data sources to ensure the model learns accurate inference patterns and retains some erroneous samples to improve the model's exploration capabilities.

Klear-Reasoner's technical principles

• Long CoT SFT (Supervised Fine-tuning)Supervised fine-tuning using high-quality data sources ensures the model learns accurate inference patterns. Prioritize a small number of high-quality data sources to avoid noise introduced by low-quality data. Retaining some erroneous samples, especially in challenging tasks, helps the model's exploratory capabilities.
• Reinforcement Learning (RL)Reinforcement learning is used to further enhance the model's reasoning ability, especially in mathematical and coding tasks. A soft reward mechanism is used, rewarding based on the pass rate of test cases to alleviate reward sparsity and improve training efficiency. Data with problematic test cases is filtered out to ensure high-quality training data.
• GPPO (Gradient-Preserving Clipping Policy Optimization) algorithmIn traditional PPO and GRPO algorithms, the clip operation discards the gradients of high-entropy tokens, limiting the model's exploration capabilities and causing delayed convergence for negative samples. GPPO, based on the stop gradient operation, decouples the clip operation from backpropagation of gradients, preserving the gradients of all tokens. For high-entropy tokens, GPPO preserves their gradients and constrains them within a certain range. For negative sample tokens, GPPO preserves the gradients and constrains them within a certain amplitude, accelerating error correction.
• Soft reward mechanismIn reinforcement learning for code tasks, soft reward mechanisms (rewarding based on the pass rate of test cases) are more effective than hard rewards (scores for complete pass, zero for failure). Soft reward mechanisms can alleviate the reward sparsity problem, increase the density of training signals, reduce the variance of gradient estimation, and make the model's learning process more stable and efficient.

Klear-Reasoner project address

• GitHub repositoryhttps://github.com/suu990901/KlearReasoner/
• HuggingFace model libraryhttps://huggingface.co/Suu/Klear-Reasoner-8B
• arXiv technical paper: https://arxiv.org/pdf/2508.07629

Application scenarios of Klear-Reasoner

• EducationAs an intelligent math tutor, it provides students with detailed problem-solving steps and reasoning processes to help users better understand and master mathematical knowledge.
• Software developmentIt automatically generates high-quality code snippets to help developers quickly implement functional modules, while also providing code review suggestions to improve code quality and development efficiency.
• FintechAnalyzing financial data to conduct risk assessments and forecasts provides logical reasoning support for financial institutions, facilitating more accurate decision-making.
• Scientific Research and Data AnalysisKlear-Reasoner can handle complex data analysis and scientific computing tasks, providing researchers with logical reasoning and model interpretation to improve research efficiency.
• Intelligent Customer ServiceIt can quickly and accurately answer complex user questions, provide a clear reasoning process, and improve user experience and problem-solving efficiency.