DeepSeek-R1, a new AI model (模型) developed by the Chinese company DeepSeek, has rocked the AI industry around the world. Open to the public on Jan. 20, it quickly beat OpenAI’s ChatGPT and topped the Apple’s App Store by Jan. 27.
According to DeepSeek, in tasks like math and natural language reasoning, the performance of this model matches that of the leading models from big companies like OpenAI. But it’s much cheaper and uses much less computing power. The report on its earlier V3 model showed that DeepSeek is the least expensive among large language models. It costs about 5.57 million US dollars (about 40.58 million yuan) to make it.
Behind the lower cost is the introduction of a new idea in R1’s training. Different from traditional methods such as CoT (思维链) and SFT (监督微调), DeepSeek uses RL (强化学习) as the main training method. While CoT depends on step-by-step reasoning and SFT on huge amounts of data (数据), R1 allows developing reasoning skills in a natural way, making it better suited for difficult and changing tasks.
The AI industry’s development has long depended on making computing power bigger. “By increasing the data quality and improving the model structure, DeepSeek may change the AI field,” pointed out the US bank Morgan Stanley. “Bigger is no longer always smarter,” said the bank.
What’s more, DeepSeek-R1 is open-source. Everyone is free to the model’s code (代码) and makes changes to suit their own needs. “We won’t choose closed-source,” said Liang Wenfeng, the founder of DeepSeek. He highlights the importance of sharing knowledge, building a healthy field and helping the progress of technology. “That is the power of open research,” Meta’s chief AI scientist Yann LeCun agreed.