Meta has publicly released Llama 3.1, introducing its flagship 405B parameter model alongside upgraded 8B and 70B versions. The company positions the 405B model as the world's largest openly available foundation model, claiming it rivals top closed AI models in capabilities such as general knowledge, math, and tool use.
- Llama 3.1 405B was trained on over 15 trillion tokens using more than 16,000 H100 GPUs and quantized to 8-bit (FP8) numerics for efficient inference.
- The upgraded 8B and 70B models feature a 128K context length, multilingual support, and improved reasoning capabilities.
- Meta updated the license to allow developers to use model outputs to improve other models, including through synthetic data generation and distillation.
- Performance evaluations suggest competitiveness with GPT-4, GPT-4o, and Claude 3.5 Sonnet across over 150 benchmark datasets.
- The release includes Llama Guard 3, Prompt Guard, and a proposal for "Llama Stack" to standardize component interfaces.
The release aims to lower barriers for developers by providing customizable weights that can be run on-premises or locally without sharing data with Meta.