Kuaishou Company officially released and open-sourced the KAT-V1 AutoThink large model. This model demonstrates excellent performance in integrating thinking and non-thinking capabilities, and can automatically adjust its thinking mode based on the complexity of the question.
KAT-V1 has two versions: 40B and 200B. The 40B version performs exceptionally well, with performance close to the recently released DeepSeek-R1 (with 685 billion parameters) in auto-think mode. The 200B version outperforms the flagship models of Qwen, DeepSeek, and Llama series in multiple benchmark tests.
Significant Performance Advantages
In the real-time benchmark test LiveCodeBench Pro, the 40B version of KAT-V1 successfully entered the closed-source model group, surpassing many open-source models.
The Kwaipilot team at Kuaishou detailed multiple technological innovations behind KAT-V1 in their technical report, including a new hybrid training paradigm for short and long thinking, as well as a novel reinforcement learning algorithm called Step-SRPO, which significantly improves the model's reasoning ability and thinking density.
Addressing the Issue of Overthinking
Since OpenAI launched the o series model, the thinking style of reasoning models has gradually evolved into "overthinking," leading to longer response times and a decline in user experience.
KAT-V1 is optimized to address this issue. The research team hopes that the model can autonomously decide whether to engage in deep thinking based on the task complexity, thus achieving more efficient human-computer collaboration. Kwaipilot's KwaiCoder-AutoThink-preview released in June this year provided an initial solution to this problem, and KAT-V1 further optimized its reasoning capabilities based on this.
Innovative Data Processing and Model Distillation Technologies
The KAT-V1 model is an extension of Qwen2.5-32B. The team constructed a large amount of thinking and non-thinking data, and used approximately 10 million examples during the pre-training phase to ensure the model's capability generalization across multiple fields such as science, code, and mathematics. Through a unique heterogeneous distillation framework, KAT-V1 efficiently transfers knowledge from the teacher model to the student model, greatly reducing the cost of model initialization.
In the post-training phase, the Kwaipilot team improved the model's intelligent decision-making ability through reinforcement learning methods. KAT-V1 can learn to intelligently choose the appropriate thinking mode, enabling its performance on complex problems to reach over 95% of DeepSeek-R1-0528.
Currently, the 40B version of KAT-V1 is available on the Hugging Face platform, and users can also experience this model through Kwaipilot, the AI development assistant created by Kuaishou. The 200B version of the MoE model is still under training, and it is expected to bring even stronger features and applications in the future.
Model Open Source Address: https://huggingface.co/Kwaipilot/KAT-V1-40B
Technical Report Address: https://arxiv.org/pdf/2507.08297