The company said it has initially released the first two models of the current version, featuring 8B and 70B parameters, with upcoming models slated to feature 400B parameters.

Meta emphasized Llama 3  was trained with a “large, high-quality training dataset” featuring over 15 trillion tokens, 7x larger than Llama 2 and featuring 4x more code. Llama 3 also features filtering techniques, including NSFW filters, to ensure data quality.

LLama 3 outperforms Llama 2 and competing models such as Anthropic’s Claude Sonnet, Mistral Medium, and OpenAI’s Chat GPT-3.5 more than half the time across 12 use cases.

The first releases of Llama 3 are text-based models. However, future releases will be multilingual and multimodal. They will also feature longer context and demonstrate better performance in reasoning and coding, which Meta described as “core LLM capabilities.”

The company plans to deploy Llama 3 on all major cloud providers, model API providers, and other services. It plans to launch the product “everywhere.”

Broader user access

Llama 3 is aimed at developers, but Meta has also launched new ways for end users to access AI services in the US and more than 12 other countries.

One new addition is a dedicated website called Meta AI, where users can access AI-powered writing assistance, trivia games, simulated job interviews, and homework help.

Meta has also integrated Meta AI with all its products, including Facebook, Instagram, WhatsApp, and Messenger. Furthermore, the company offers the service via Ray-Ban Meta smart glasses in the US and plans to extend it to its Meta Quest VR headset.

In an announcement, CEO Mark Zuckerberg said:

“We believe that Meta AI is now the most intelligent AI assistant that you can freely use.”

News of Meta’s expanded AI products comes shortly after upgrades to competing services. ChatGPT upgraded to GPT-4 Turbo on April 11, while Microsoft Copilot upgraded to GPT-4 Turbo starting in March, advancing the race between consumer-oriented AI services.

Mentioned in this article