Announced in October this yearOpen Source Large Natural Language Model Llama 3Later, Meta announced the launch ofNew version Llama 3.1The maximum scale can correspond to 4050 billion sets of parameters, and is also provided to the outside world in an open source form.
Llama 3.1 was trained on over 16000 NVIDIA H100 accelerators and a dataset of over 15 trillion tokens, though the specific source of the dataset was not disclosed. With the release of Llama 3.1, Meta also upgraded the 3 billion and 700 billion parameter versions of Llama 80.
Llama 3.1 is also available as open source and will be used on service platforms such as AWS, Google Cloud, Azure, NVIDIA NIM, Snowflake, Databricks, Kaggle, and IBM WatsonX. It will also be available through the Hugging Face platform.
According to Meta, Llama 3.1 surpasses OpenAI's GPT-4o and Anthropic's Claude 3.5 Sonnet in many test projects, and the cost of running Llama 3.1 on AWS cloud services is only about half of that of GPT-4o.
Meta CEO Mark ZuckbergExplanationThe reason why the Llama artificial intelligence model is made available in an open source format is that it hopes to attract more people to use this artificial intelligence model to create various application services by emulating the open source model of the Linux operating system. This will allow the Llama artificial intelligence model to continue to improve and become a technology used by many industries.
Meta also announced that its artificial intelligence assistant service Meta AI will be expanded to support 22 languages, mainly including Argentinian, Chilean Spanish, Colombian Spanish, Mexican, Peruvian Spanish, and Cameroonian Pidgin. It will also be available on the Quest series of virtual vision headsets in the US and Canadian markets. In addition, it also launched a profile photo automatic image generation function called Imagine me, allowing users to create more distinctive personal representative images.
