“The amount of compute needed to train Llama 4 will likely be almost 10x more than what we used to train Llama 3, and future models will continue to grow beyond that.”

Mark ZuckerbergCEO

Founder, Chairman and CEO of Meta Platforms

Llama 4 as actually trained fell far short of a 10x compute scale-up over Llama 3.1 405B (~3.8e25 FLOP): Epoch AI estimates Llama 4 Maverick at roughly 2.2e24 FLOP, and the largest model, Behemoth — reported as trained on about 32,000 H100s — at roughly 5e25 FLOP, only around 1.3x Llama 3.1 405B. Behemoth was delayed through 2025 and never publicly released.