Meta has just released its largest and most advanced artificial intelligence (AI) model. We are talking about Llama 3.1, an “open source” proposal that directly rivals OpenAI’s GPT-4o and Anthropic’s Claude 3.5 Sonnet, beating them in several benchmarks commonly used by the industry to evaluate the performance and efficiency of these technologies.

The social media company’s new flagship model lands on Meta AI, the AI ​​assistant that is present on WhatsApp, Facebook, and Instagram in some markets, but which still does not have a date of arrival in the countries that make up the European Union (EU). However, those with compatible hardware can download it from anywhere in the world to use it locally.
Outstanding in mathematics, translation, and more

As we know, not all models are perfect for everything. In this case, we are faced with an alternative that, according to Meta, excels in general knowledge, mathematics, tool use, and multilingual translation. As we can see in the image below, Llama 3.1 405B (405 billion parameters) stands out in several aspects compared to its main competitors.

But Llama 3.1 is not just this huge language model. The company has also launched smaller versions that can be very useful for a more specific range of tasks. Llama 3.1 70B is aimed at content creation, conversational AI, language understanding, and business applications. It can do everything from summarizing text to generating programming code.

Llama 3.1 8B is the smallest of the alternatives. With 8 billion parameters, it can run on less powerful computers. For its part, it also promises to be very useful in analyzing and classifying texts. Here, the speed of the responses stands out. We don't know how many parameters OpenAI's new GPT-4o mini has, but they could probably compete in the same category of solutions.
The open approach of Llama 3.1

The company led by Mark Zuckerberg has been working on “open source” models for some time, although not everyone agrees with the use of this term. Some point out that the license of these models has limitations that do not allow them to be classified in this way. Llama 3.1, in fact, makes use of “open source” again, with new features at the license level.

The company's latest model is available for free use, even for commercial purposes, unless it is exposed to hundreds of millions of users. Zuckerberg, in fact, has published a letter in which he has pointed out that “open source AI is the way forward” and that open source models are increasingly better and will end up surpassing closed models at some point.

Having the possibility of downloading a model of these characteristics to use it with a certain freedom is no small thing. The preparation of Llama 3.1 405B has required several months of training with 16,000 NVIDIA GPU H100 graphics cards. Meta has also released the model weights so that companies can train it with custom data and adjust it to their own needs.

Beyond the fact that the variants of this model can run locally, we live in a world where cloud computing is a reality. Moving a colossus of 405 billion parameters is not something that can be done with any hardware. Here, precisely, business clients will be able to hire various services, such as AWS, Google Cloud, Databricks or Microsoft Azure.
What is happening with Meta in the EU?

We are witnessing the intention of the EU to impose strict regulations on the AI ​​industry. What is happening is that many companies are delaying their launches. Meta, the protagonist of this article, announced a few days ago that it would not launch its next multimodal model in the community territory due to the unpredictable nature of the European regulatory environment.”

Llama 3.1 is now available globally, meaning users in EU member states will be able to use it, the company said. They won't be able to access it through certain tools that, as we noted above, have not yet officially landed in the EU, such as Meta AI. We'll have to wait and see how the launches of AI products in the EU will evolve.

https://www.xataka.com/robotica-e-ia/meta-acaba-lanzar-modelo-ia-grande-capaz-mundo-supera-varios-aspectos-a-gpt-4o-openai