DeepSeek, a Chinese AI company, released DeepSeek-R1, an open-source reasoning model, stating that this model has surpassed OpenAI’s o1 model on key performance benchmarks. Earlier, DeepSeek, the Hangzhou-based company had unveiled the DeepSeek V3 model and claimed that it outperformed Meta’s Llama 3.1 and OpenAI’s GPT-4o.
Designed for advanced problem-solving and analytical functions, DeepSeek-R1 consists of two core versions: DeepSeek-R1-Zero and DeepSeek-R1. The DeepSeek-R1-Zero is trained through the reinforcement learning (RL) method without any supervised fine-tuning. On the other hand, DeepSeek-R1 is built on DeepSeek-R1-Zero with a cold-start phase, efficiently curated data, and multi-stage RL.
According to the technical report released by DeepSeek, DeepSeek-R1 has performed well on several important benchmarks. It scored 79.8 percent (Pass@1) on the American Invitational Mathematics Examination (AIME) 2024, slightly surpassing OpenAI’s o1. DeepSeek-R1 also achieved an accuracy of 93 percent on the MATH-500 test.
Read More: OpenAI to Introduce PhD Level AI Super-Agents: Reports
Demonstrating its coding capabilities, DeepSeek secured a 2029 Elo rating on the Codeforces and performed better than 96.3 percent of human participants. It scored 90.8 percent and 71.5 percent on the general knowledge benchmarks MMLU and GPQA Diamond, respectively. To test writing and question-answering capabilities, DeepSeek-R1 was tested on the AlpacaEval 2.0 benchmark and achieved an 87.6 win rate.
Such high-performance caliber makes DeepSeek-R1 suitable for solving complex mathematical problems and code generation in software development. Its ability to generate responses in a stepwise manner, like human reasoning, makes DeepSeek-R1 useful for research, attracting the attention of the scientific community.
Launched under the open-source MIT license, DeepSeek-R1 can be freely used by enterprises for commercial purposes. However, they will have to spend an additional amount on customization and fine-tuning. In addition, companies outside China may be skeptical about using DeepSeek-R1 due to AI regulatory challenges and geopolitical reasons.