DeepSeek-R1, a large language model developed by a Chinese startup, has emerged as an affordable and open rival to models like OpenAI’s o1. Unlike its counterparts, R1 generates responses step-by-step, mimicking human reasoning, making it more effective in solving complex scientific problems. Initial tests showed its performance on tasks such as chemistry, mathematics, and coding was comparable to that of o1, which had recently been released. DeepSeek has even made the model’s training data publicly available, a rare feature for AI models built by major companies like OpenAI.
R1 is priced at one-thirtieth the cost of running similar tasks with o1, making it highly cost-effective. To accommodate researchers with limited resources, DeepSeek has also developed simplified versions of R1, referred to as “distilled” models. For example, a task that would have previously cost over £300 using o1 can now be completed for less than $10 with R1. This affordability and enhanced reasoning capabilities are significant steps forward in the race between Chinese and Western AI firms.
The model’s success also highlights the growing competition among LLMs (large language models) and challenges existing paradigms, particularly regarding access to advanced computing resources despite US export controls. Critics note that this development signals a narrowing of the lead once held by the United States, suggesting a potential shift toward collaborative AI advancements rather than an arms race in innovation.
Source: https://www.nature.com/articles/d41586-025-00229-6