Meta Meta on Tuesday announced the latest version of its Llama AI model, dubbed Llama 3.1. The latest Llama comes in three different versions, with one variant being Meta’s largest and most efficient AI model yet. Like previous versions of Llama, the latest model remains open source, meaning it’s free to access.
The new large language model, or LLM, underscores the social network's massive investment to keep pace with AI spending from the likes of leading startups OpenAI and Anthropic and other tech giants like Google and Amazon.
The announcement also highlights the growing partnership between Meta and NvidiaNvidia is a key partner of Meta, providing the Facebook parent company with computing chips called GPUs to help train its AI models, including the latest version of Llama.
While companies like OpenAI aim to make money by selling access to their law master's programs or offering services to help customers use the technology, Meta has no plans to start its own competing business, a Meta spokesperson said during a media briefing.
Instead, much like when Meta launched Llama 2 last summer, the company is partnering with a handful of tech companies that will provide customers with access to Llama 3.1 via their cloud computing platforms, as well as sell security and management tools that work with the new software. Meta’s 25 Llama-related partners include Amazon Web Services, Google Cloud, Microsoft Azure, Databricks, and Dell.
While Meta CEO Mark Zuckerberg has told analysts on past earnings calls that the company generates some revenue from its partnerships with Lama, a Meta spokesperson said any financial benefit is merely incremental. Instead, Meta believes that by investing in Lama and related AI technologies and making them freely available through open source, it can attract high-quality talent in a competitive market and lower overall computing infrastructure costs, among other benefits.
Meta's launch of Llama 3.1 comes ahead of a conference on advanced computer graphics, where Zuckerberg and Nvidia CEO Jensen Huang are scheduled to speak together.
The social media giant is one of Nvidia’s largest customers that doesn’t run its own commercial cloud, and Meta needs the latest chips to train its AI models, which it uses internally for targeting and other products. For example, Meta said the largest version of its Llama 3.1 model announced Tuesday was trained on 16,000 Nvidia H100 GPUs.
But the relationship is also important to both companies because of what it represents.
For Nvidia, the fact that Meta is training open-source models that other companies can use and adapt for their own business — without paying licensing fees or asking for permission — could expand the use of Nvidia's chips and keep demand high.
But open-source models can cost hundreds of millions or billions of dollars to build. And there aren’t many companies that have the financial capacity to develop and launch such models with similar amounts of investment. And while Google and OpenAI are Nvidia customers, they keep their most advanced models available to everyone.
On the other hand, Meta needs a reliable source of the latest GPUs to train increasingly powerful models. And like Nvidia, Meta is trying to foster an ecosystem of developers who build AI applications with the company’s open-source software at the core, even if Meta has to give up code and so-called AI weights that are expensive to build.
The open-source approach benefits Meta by exposing developers to its internal tools and inviting them to build on them, Ash Jhaveri, the company’s vice president of AI partnerships, told CNBC. It also helps Meta because it uses its AI models internally, allowing the company to reap the improvements made by the open-source community, he added.
Zuckerberg wrote in a blog post on Tuesday that he is taking a “different approach” to Llama’s release this week, adding: “We are actively building partnerships so that more companies in the ecosystem can offer unique functionality to their customers as well.”
Since Meta is not an institutional vendor, it can refer companies inquiring about Llama to one of its institutional partners, such as Nvidia, Jhaveri said.
The largest version of the Llama 3.1 model family is called Llama 3.1 405B. This model has 405 billion parameters, which refers to the variables that determine the overall size of the model and the amount of data it can process.
In general, a large LLM program with a large amount of parameters can perform more complex tasks than smaller LLM programs, such as understanding context in long text streams, solving complex math equations, and even generating synthetic data that can be used to improve smaller AI models.
Meta is also releasing smaller versions of Llama 3.1 called Llama 3.1 8B and Llama 3.1 70B. The company said these versions are essentially upgraded versions of previous versions and can be used to run chatbots and coding assistants.
Meta also said that the company’s WhatsApp users in the US and visitors to its Meta.AI website will be able to see Llama 3.1’s capabilities by interacting with the company’s digital assistant. The digital assistant, which will run on the latest version of Llama, will be able to answer complex math problems or solve programming problems, a Meta spokesperson said.
WhatsApp and Meta.AI users in the US will be able to switch between the new giant Llama 3.1 LLM or a less capable but faster and smaller version to get answers to their queries, a Meta spokesperson said.
