Meta 推出 Llama 3.1：405B、70B 與 8B 旗艦開源模型，支援多語言與 128K 超長上下文

Meta's Llama 3.1 represents a major milestone in the open-source AI landscape. The most notable model is the 405B (405 billion parameter) version — the first open-source model capable of standing toe-to-toe with top-tier commercial models such as GPT-4o and Claude 3.5 Sonnet across domains including commonsense reasoning, mathematics, and translation. Llama 3.1 also updates the existing 8B and 70B models. Three major highlights of this upgrade are: a 128k extended context window (a dramatic improvement over the previous 8k limit), robust multilingual support (officially supporting 8 languages), and significantly enhanced tool use capabilities. Crucially, Meta revised its license terms to allow developers to use outputs from Llama 3.1 to train and improve other models, which will greatly accelerate distillation and innovation within the open-source community.