A large language model developed in China, known as DeepSeek-R1, is exciting scientists as a cost-effective and open alternative to reasoning models like OpenAI's o1. These models produce answers in a sequence that mimics human reasoning, making them more skilled than older language models in tackling scientific issues, potentially proving beneficial for research. Early assessments of R1, which launched on January 20, indicate that its performance in tasks related to chemistry, mathematics, and coding rivals that of o1— which impressed researchers upon its release by OpenAI in September. “This is astonishing and completely unexpected, ” commented Elvis Saravia, an AI researcher and co-founder of the UK-based AI consulting firm DAIR. AI, on X. R1 is notable for another reason. The start-up DeepSeek, based in Hangzhou, has made the model ‘open-weight’, allowing researchers to analyze and enhance the algorithm. It’s published under an MIT license, enabling free reuse, though it’s not fully open source, as the training data remains undisclosed. “The transparency shown by DeepSeek is remarkable, ” states Mario Krenn, head of the Artificial Scientist Lab at the Max Planck Institute for the Science of Light in Erlangen, Germany. In contrast, o1 and other models from OpenAI in San Francisco, including the recent o3, are considered “essentially black boxes, ” he notes. While AI hallucinations cannot be entirely eliminated, certain techniques can help mitigate their impact. DeepSeek hasn't disclosed the total cost of training R1, but it charges users about one-thirtieth of o1’s operational expenses. The company has also developed mini ‘distilled’ versions of R1, allowing researchers with limited computing resources to experiment with the model. Krenn notes, “An experiment that cost over £300 with o1 was less than $10 with R1, ” highlighting a significant difference that could influence future adoption. R1 is part of a surge in Chinese large language models (LLMs).

Emerging from a hedge fund, DeepSeek gained recognition last month with the release of a chatbot named V3, which surpassed major competitors despite a limited budget. Experts estimate that the hardware rental for training the model was around $6 million, compared to over $60 million for Meta's Llama 3. 1 405B, which utilized 11 times the computing resources. The excitement around DeepSeek stems in part from its achievement in developing R1 amidst U. S. export controls restricting Chinese companies' access to top-tier AI processing chips. “The fact that it originates from China demonstrates that efficient resource management is more critical than sheer computing power, ” asserts François Chollet, an AI researcher based in Seattle, Washington. DeepSeek’s advancements indicate that “the gap previously perceived as a U. S. lead has significantly narrowed, ” observed Alvin Wang Graylin, a technology expert in Bellevue, Washington, who is affiliated with the Taiwan-based immersive technology company HTC, on X. “Both countries should consider a cooperative strategy for developing advanced AI rather than continuing the current futile arms race. ”