Skip to content

Meta claims open-source Llama 3.1 is “world’s largest” model

  • by
  • 2 min read

Meta released its first openly accessible AI model, Llama 3.1, on Tuesday; and claimed that it surpasses GPT-4o and Claude 3.5 Sonnet on several benchmarks.

Llama 3.1 is more complex than the Llama 3 models released a few months earlier. The biggest version, trained on more than 15 trillion tokens using 16,000 Nvidia H100 GPUs, consists of 405 billion parameters. The company has not disclosed the cost of developing the new AI model.

The company said that the 405-billion-parameter model needed an equivalent of 30.84 million GPU hours and released 8,930 tons of CO2 emissions. Meta insisted that the said computing hours and energy were required to train the model at a large scale in a significant time.

Despite its smaller size compared to rival AI models, users still require a powerful system to run Llama smoothly. The Llama 3.1 405B needs around 810 GB of memory to work at the full 16-bit precision it was trained at. Meta released an 8-bit quantized version of the model.

Meta, in an announcement, said that its experimental evaluation suggested that their flagship model is “competitive with leading foundation models across a range of tasks, including GTP-4, GPT-4o, and Claude 3.5 Sonnet.” The model has new security and safety tools such as Llama Guard 3 and Prompt Guard.

The three new 3.1 models (8B, 70B, and 405B) have been improved with support for eight languages (English, German, French, Italian, Portuguese, Spanish, Thai, and Hindi) and a larger 1,28,000 token context window.

The company has not disclosed the new Llama model’s training data. CEO Mark Zuckerberg predicts that Meta AI will become the most widely used AI assistant by the end of the year and outperform ChatGPT. The company is also creating a Llama-based Meta AI assistant that would be accessible in many countries and languages alongside an image generation feature that would work on the basis of individual preferences.

The “Imagine Me” feature of Meta AI would scan a user’s face and allow them to insert themselves into AI-generated images. Like Meta’s integration with Ray-Ban glasses, it will be available on the Quest headset in the next few weeks.

In the News: Threat actors exploit Hamster Kombat craze to drop malware

Arun Maity

Arun Maity

Arun Maity is a journalist from Kolkata who graduated from the Asian College of Journalism. He has an avid interest in music, videogames and anime. When he's not working, you can find him practicing and recording his drum covers, watching anime or playing games. You can contact him here: arunmaity23@proton.me

>