Meta Has Unveiled the ‘World’s Largest and Most Capable’ AI Model, Surpassing OpenAI’s GPT-4o in Various Aspects

On Tuesday, Meta launched its most advanced artificial intelligence model, Llama 3.1. This “open source” model competes directly with OpenAI’s GPT-4o and Anthropic’s Claude 3.5 Sonnet, surpassing them in several industry benchmarks for evaluating performance and efficiency.

Llama 3.1, the company’s flagship model, is now available on Meta AI, its AI assistant platform available on WhatsApp, Facebook, and Instagram in some markets, including the U.S. Users with compatible hardware can download it globally for local use.

A+ in Math, Translation, and More

Not all AI models are perfect for every task. However, in this case, Meta claims Llama 3.1 is an alternative that excels in general knowledge, math, use of tools, and multilingual translation. As shown in the image below, Llama 3.1 405B (meaning 405 billion parameters) outperforms its main competitors in several aspects.

But Llama 3.1 isn’t just a large language model. The company has also launched smaller versions that can be useful for specific tasks. Llama 3.1 70B focuses on content creation, conversational AI, understanding language, and business apps. It can also perform tasks ranging from summarizing text to generating programming code.

For its part, Llama 3.18B is the smallest of the options. With 8 billion parameters, it can operate on less powerful computers. Additionally, Meta expects it to be highly effective in text analysis and classification, particularly due to its quick response times. OpenAI hasn’t provided information on the number of parameters in its new GPT-4o mini, but it will likely fall into a similar category of solutions.

Llama 3.1’s Open Source Approach

The company led by Mark Zuckerberg has been working on “open source” models for some time, although not everyone agrees with the company's use of the term. Some argue that the models’ license has limitations that prevent them from being categorized as open source. Once again, Llama 3.1 utilizes “open source,” introducing new features at the licensing level.

Meta’s latest model is available for free use, even for commercial purposes, unless companies want to use it with hundreds of millions of users. Zuckerberg has published a letter in which he states that “open source AI is the path forward” and that open source models are improving and will eventually surpass closed source models at some point.

The ability to download a model with these characteristics for flexible use is significant. Developing Llama 3.1 405B involved several months of training using 16,000 Nvidia H100 GPUs. Meta has also made the model’s weights available, allowing companies to customize their training data and tailor the model to their specific requirements.

While local variants of this model are available, it’s worth noting that cloud computing is now a reality. Moving a model with 405 billion parameters requires substantial hardware capabilities. For this reason, enterprise customers can access various services from providers such as AWS, Google Cloud, Databricks, and Microsoft Azure.

What’s Going on With Meta in the EU?

While this isn’t an issue in the U.S. at the moment, we’ve observing the EU’s plans to enforce strict regulations on the AI industry for some time now. In response, many companies are delaying their product launches. For instance, Meta recently announced that it wouldn’t introduce its next multi-modal model in the EU due to the unpredictable regulatory environment in Europe.

Although Llama 3.1 is available globally, including in the U.S., users in EU member countries won’t have access to certain tools, such as Meta AI, that haven’t yet been officially introduced in the region. It remains to be seen how the launch of AI products in the EU will develop.

This article was written by Javier Marquez and originally published in Spanish on Xataka.

Related | The EU Won’t Receive Meta’s Next Major AI Models. The Reason: Local Regulations