China’s DeepSeek Unveils AI Model Halving Costs by 50% – The ‘Sparse Attention’ Revolution
DeepSeek on Sept. 29 released DeepSeek-V3.2-Exp, an open-source language model using a new “sparse attention” method to cut API inference costs by about 50% for long texts. The model is available under an MIT license and supports Chinese AI chips. Early tests show token costs dropping from $0.07 to $0.028 per million, though performance still trails top models like GPT-5.