Meta released Llama 3.1 in three sizes — 8, 70, and 405 billion parameters — the largest a dense Transformer with a 128,000-token context window, sixteen times its predecessor, trained on more than 15 trillion tokens across 16,000 H100s. It supported tool use and eight languages and was licensed for commercial use. It was the first time a model that could hold its own against the closed frontier was available to anyone as weights. It completed the Llama 3 line begun in April with the 8B and 70B models. In an open letter the same day, Mark Zuckerberg argued that AI would follow the path from closed Unix to open-source Linux, with open models becoming the industry standard. Being able to own frontier-level capability rather than rent it fed straight back into how the other labs priced and published.