Meta said on 23 July 2024 Llama 3.1, describing Llama 3.1 405B as its first frontier-level open-source AI model, with upgraded 8B and 70B models, a 128K context length and support for eight languages.
Meta said the 405B model rivals leading closed models on general knowledge, math, tool use and multilingual translation, and that licence changes let developers use Llama outputs — including from the 405B — to improve other models.
New safety components include Llama Guard 3 for multilingual moderation and Prompt Guard against prompt injection. Meta also opened a request for comment on a Llama Stack API for toolchain and agent interfaces.
Models are available to download on llama.meta.com and Hugging Face, with day-one services from partners including AWS, NVIDIA, Databricks, Groq, Dell, Azure, Google Cloud and Snowflake.
Meta said it trained the 405B on more than 15 trillion tokens using over 16,000 H100 GPUs, and that it quantised production inference from BF16 to FP8 so the model can run within a single server node.