Performance benchmark reports from reputable AI research publications or organizations

DeepSeek-R1 Matches OpenAI's o1 at 98% Lower Cost, Boosts Open-Source AI Innovation

Jan 26, 2025, 03:33 AM

DeepSeek, a Chinese AI research company, has released DeepSeek-R1, an open-source AI model that matches or exceeds the performance of OpenAI's o1 model at a cost that is 98% lower. DeepSeek-R1 was developed using a unique training process that emphasizes reinforcement learning (RL) over traditional supervised fine-tuning (SFT), allowing it to achieve advanced reasoning capabilities. The model utilizes a Mixture of Experts design with 671 billion parameters, activating only 37 billion parameters per forward pass. It employs Group Relative Policy Optimization (GRPO) and integrates cold-start data, reasoning-oriented RL, and SFT using a dataset of approximately 800,000 samples. DeepSeek-R1 has demonstrated strong performance in mathematics, achieving a Pass@1 score of 97.3% on the MATH-500 benchmark, and in coding, with a Codeforces Elo rating of 2029. It also scored 79.8% on the AIME 2024 benchmark, showcasing its advanced reasoning capabilities. This development challenges the dominance of proprietary AI models and highlights the potential of open-source innovation in the AI industry.