"Llama 4 Maverick needed 10 times more compute to match the same quality. DeepSeek needed eight times more compute and Kimi needed three times more compute to match the same quality."