KAT-V1 - Kuaishou's open-source automated thinking model

What is KAT-V1?

KAT-V1 is Kuaishou's open-source AutoThink model, available in 40B and 200B versions. The model integrates thinking and non-thinking abilities, automatically switching thinking modes based on question difficulty. Version 40B's performance approaches that of DeepSeek-R1 (685 billion parameters), while version 200B surpasses open-source models such as Qwen, DeepSeek, and Llama in multiple benchmark tests. KAT-V1 uses a hybrid long and short thinking training paradigm and the novel reinforcement learning method Step-SRPO to enhance thinking density and judgment, efficiently completing cold starts through a heterogeneous distillation framework. The model performs exceptionally well in complex inference tasks such as code generation and SQL optimization, and supports user-guided thinking modes.

Main functions of KAT-V1

• Switching between automatic thinking and non-thinking modesThe system automatically determines whether to enter a thinking mode based on the complexity of the problem. For complex problems, the model initiates deep reasoning and planning; for simple problems, it directly provides the answer, avoiding unnecessary waste of computational resources.
• Complex reasoning abilityWhen faced with complex programming tasks (such as generating code to simulate a ball moving within a rotating hexagon) and SQL optimization tasks, KAT-V1 provides structured, multi-step analysis and solutions.
• Multi-turn dialogue capabilityKAT-V1 supports multi-turn dialogues, allowing solutions to be gradually improved based on user needs.
• User Intent GuidanceThe model allows users to guide the model to activate thinking mode using simple intention commands (such as explicit thinking or non-thinking preferences).
• Intelligent agent mode adaptation: Adapt to multi-agent scenarios, such as disabling inference during document inspection and enabling deep inference when diagnostics or code generation are required.

KAT-V1 Technical Principles

• Training Paradigm of Hybrid Long-Term and Short-Term ThinkingKAT-V1 introduces a new hybrid long-short thinking model training paradigm, combining the traditional reinforcement learning algorithm (GRPO) with the novel reinforcement learning method Step-SRPO. This paradigm improves the thinking density of the model's output tokens and its ability to determine whether to activate the thinking mode.
• Heterogeneous distillation frameworkKAT-V1 uses a unique heterogeneous distillation framework, comprising two main modules: Universal Logits Distillation Loss (ULD Loss) and Multi-Token Prediction (MTP). In this way, the model can complete the cold start at a lower cost while improving the efficiency of knowledge transfer.
• Step-SRPO reinforcement learning algorithmThe Step-SRPO algorithm guides model learning through a dual reward mechanism (judgment reward and answer reward), allowing the model to gradually learn to flexibly adjust the inference depth according to the difficulty of the question during training, thereby further reducing the use of tokens while improving model performance.
• High-quality data synthesisDuring the pre-training phase, a large amount of high-quality thinking/non-thinking data is used. The thinking data is synthesized through the Agentic framework, which consists of solvers, thinkers, and commenters, ensuring the logical consistency of the synthesized data and the quality of the output.

KAT-V1 Project Address

• HuggingFace model libraryhttps://huggingface.co/Kwaipilot/KAT-V1-40B
• arXiv technical paper: https://arxiv.org/pdf/2507.08297

Application scenarios of KAT-V1

• Code generation and optimizationKAT-V1 can generate complex code, such as Python code simulating the movement of a ball within a rotating hexagon, and provides SQL optimization suggestions.
• Complex Reasoning and Problem SolvingIt automatically assesses the difficulty of the problem, initiates a deep reasoning mode, and supports multi-turn dialogue to gradually improve the solution.
• Multi-agent scenariosThe model is adapted to multi-agent scenarios, supporting agent collaboration and task allocation, such as file inspection and code generation.
• User Intent GuidanceIt allows users to guide the model to activate its thinking mode through simple commands, providing personalized services.
• Multimodal and interactive applicationsIt is expected to be expanded to multimodal applications in the future, supporting real-time interaction and dynamic adjustment.