“The DeepSeek v3 AI model achieved results comparable to Meta's Llama 3.1 using fewer than 2.8 million training hours, a roughly 10x improvement over Llama's nearly 31 million hours.”