DeepSeek-AI introduces its first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1, aimed at improving reasoning capabilities in large language models (LLMs) through reinforcement learning (RL). DeepSeek-R1-Zero, trained solely via RL, shows strong reasoning abilities but struggles with readability and language mixing. To address these issues, DeepSeek-R1 employs a multi-stage training process, including cold-start data before RL, achieving performance comparable to OpenAI's o1-1217 in reasoning tasks. The research also involves open-sourcing these models and six distilled versions based on Qwen and Llama, highlighting the importance of post-training in enhancing reasoning tasks, aligning with social values, and adapting to user preferences with minimal computational resources. Despite advancements, effective test-time scaling remains a challenge. The study uses DeepSeek-V3-Base as the base model and the GRPO framework to improve reasoning performance, with DeepSeek-R1-Zero significantly increasing the pass@1 score on AIME 2024. The distillation process from DeepSeek-R1 to smaller models, like Qwen2.5-32B, shows that smaller models can inherit reasoning patterns from larger models, leading to superior performance.