On September 18, 2025, Reuters reported that DeepSea said that training its successful model only cost about $294,000, a low figure compared to the norm. This raises a discussion about the effectiveness of training architectures and approaches, and the potential for smaller companies to be able to enter the model race. For users and the market, competition may increase and good models may emerge at a lower cost. Scientifically, this encourages efficiency research in data and computing. Future impact: Pressure to reduce the cost of model training without sacrificing quality.
China's DeepSec unveils low-cost training for a prominent model.
