Press releases or official announcements from DeepSeek or partnering companies

DeepSeek Launches 671B-Parameter DeepSeek-R1 Model, Matching OpenAI's o1 at 90-95% Lower Cost

Jan 20, 2025, 02:32 PM

DeepSeek, a Chinese AI lab, has released DeepSeek-R1, an open-source reasoning model that claims to match the performance of OpenAI's o1 across various benchmarks in math, coding, and reasoning tasks. DeepSeek-R1, along with its predecessor DeepSeek-R1-Zero, is available under an MIT license on Hugging Face, allowing for commercial use and further development. The model, which contains 671 billion parameters, was developed using large-scale reinforcement learning (RL) without initial supervised fine-tuning (SFT), a method that has shown promising results in enhancing reasoning capabilities. DeepSeek has also introduced six smaller distilled models based on DeepSeek-R1, which are derived from Llama and Qwen architectures. These distilled models, ranging from 1.5 billion to 70 billion parameters, offer competitive performance, with the 32B model outperforming OpenAI's o1-mini on several benchmarks. DeepSeek-R1 is accessible via an API, with pricing 90-95% cheaper than that of OpenAI's o1, making it an attractive option for developers and organizations.