DeepSeek announced the official release and open source of its large language model R1, whose performance is comparable to the official version of OpenAI's o1. R1 makes extensive use of reinforcement learning technology in the post-training phase. Even if the amount of data annotation is small, it can maintain strong reasoning capabilities and perform well in mathematics, code, and natural language reasoning. DeepSeek also open sourced two 660B parameter-scale models, R1 and R1-Zero, as well as 6 small models using model distillation technology, including 32B and 70B parameter models, whose performance surpassed OpenAI's o1-mini. In addition, DeepSeek also provides a highly competitive API pricing plan and adopts an MIT license, allowing unlimited commercial use and encouraging users to conduct secondary development based on R1.
DeepSeek R1 has demonstrated its strong strength and open attitude in terms of performance, open source strategy and API pricing, which marks a significant progress in domestic AI technology and provides more choices and possibilities for global AI developers. The opening of open source R1 and related resources will promote the popularization and development of AI technology and accelerate the arrival of the era of equal rights in AI technology. Visit the link for more information: Paper: https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf API documentation: https://api-docs.deepseek.com/zh-cn /guides/reasoning_model