Comment by gmd63

DeepSeek was trained with distillation. Any accurate estimate of training costs should include the training costs of the model that it was distilling.