On April 18, Meta officially released its much-anticipated open-source large language model, Llama 3.

Meta CEO Mark Zuckerberg announced that the company's AI assistant, powered by the latest Llama 3 model, is now available across all of its apps, including Instagram, WhatsApp, and Facebook. It also has its own dedicated website. Additionally, an image generator has been launched that can generate images based on natural language prompts.

In a video, Zuckerberg said that the assistant can answer questions, create animations, and generate images.

Meta CEO Mark Zuckerberg stated in a video: “We believe that Meta AI is now the most intelligent artificial intelligence assistant available for free use. Meta AI is integrated into the search boxes of WhatsApp, Instagram, Facebook, and Messenger apps, allowing users to easily ask questions that can be answered by the new tool.”

Zuckerberg mentioned that the generative AI capabilities provided by Meta are the most powerful in free products. Users can now search with Meta AI on Facebook, Instagram, WhatsApp, and Messenger without switching between apps:

When browsing the feed, you can also directly ask Meta AI for more information from the posts:

The image generator brings a more fun experience, with the Imagine feature creating images from text in real-time.

The beta version of this feature is launching today on WhatsApp in the US and on the Meta AI web experience. As you start typing, you will see an image appear, and the image will change with each additional letter typed:

Meta indicates that Llama 3 outperforms industry-leading models in multiple key benchmark tests, achieving comprehensive leadership in tasks such as code generation. It is capable of complex reasoning, better adherence to instructions, visualizing ideas, and solving many subtle problems.

Key Highlights of Llama 3:

  1. Trained on over 15T tokens, 7 times more than Llama 2's dataset

  2. Supports 8K long text, with an improved tokenizer with a vocabulary of 128K tokens for better performance

  3. State-of-the-art performance on a large number of important benchmarks

  4. New capabilities, including enhanced reasoning and coding

  5. 3 times more training efficient than Llama 2

  6. New trust and safety tools with Llama Guard 2, Code Shield, and CyberSec Eval 2

  7. The newly released 8B and 70B parameter Llama 3 models are already being used in the Meta AI assistant and are also open source for developers, including pre-trained and fine-tuned versions.

Significant Leap from Llama 2

The newly released 8B and 70B parameter Llama 3 models represent a significant leap from Llama 2. Due to improvements in pre-training and post-training, the pre-trained and instruction fine-tuned models released this time are the best 8B and 70B parameter models available today. At the same time, improvements in the post-training process have significantly reduced the model's error rate, further improved consistency, and increased the diversity of the model's responses.

Editor: Alexander