On Tuesday, the 23rd, Meta revealed the latest version of its artificial intelligence model, Llama 3.1, which comes in three variations. One of these versions stands out as Meta’s most advanced AI model to date.
Following the tradition of its previous versions, Llama 3.1 is open source, providing free access to the public.
This launch reflects Meta’s significant investment in AI, aiming to compete with startups like OpenAI and Anthropic, as well as tech giants like Google and Amazon.
Although the company has not disclosed the development costs for Llama 3.1, Meta CEO Mark Zuckerberg recently stated that the company is allocating billions to advance AI.
With this announcement, Meta demonstrates that an open approach can be as effective as the closed approach commonly adopted by other AI companies. However, this strategy places the company at the center of a debate over the risks of releasing AI without stringent controls. Meta trains Llama to avoid generating harmful content, but the model can be adjusted to remove these protections.
According to Meta, Llama 3.1 is comparable to the best commercial AI offerings, such as those from OpenAI, Google, and Anthropic. In certain AI progress benchmarks, the company claims its model is the most intelligent in the world.
Transformation in the Landscape
In an open letter accompanying the launch, Zuckerberg compared Llama to the open-source operating system Linux.
When Linux gained traction in the 90s and early 2000s, many major tech companies were investing in closed alternatives, criticizing open-source software as risky and unreliable. Today, Linux is widely used in cloud computing and serves as the basis for the Android mobile operating system.
“I believe AI will follow a similar path,” Zuckerberg wrote. “Today, several companies are developing leading closed models, but open source is rapidly narrowing that gap.”
Llama 3.1 has 405 billion parameters, or adjustable elements. Meta has also released two smaller versions of Llama 3, one with 70 billion parameters and another with 8 billion. Updated versions of these models, named Llama 3.1, were also launched today.
Although too large to run on a typical computer, Meta assures that many cloud providers, such as Databricks, Groq, AWS, and Google Cloud, will offer hosting options for developers to use customized versions of the model. Llama 3.1 is also accessible on Meta.ai.
Unlike the latest models from OpenAI and Google, Llama is not “multimodal,” meaning it was not designed to handle images, audio, and video. However, Meta claims that the model is significantly better at using other software, such as web browsers, which many researchers and companies believe can make AI more useful.
