Meta Platforms has released the latest version of the Llama 3 artificial-intelligence models, which are mostly free. The new model boasts multilingual abilities and general performance metrics to compete with paid models like OpenAI. Facebook’s parent company announced the new Llama 3 AI model in blog posts, a research paper, and other publications. The Llama 3 can speak eight languages and write better computer code. It also solves more complex math problems.
Details about the new Meta release
It is the most advanced version yet, with 405 billion variables or parameters that the algorithm considers when generating responses to queries from users. However, it still falls short of the leading models offered by its competitors.
Amazon’s GPT-4 is said to have 2 trillion parameters, while OpenAI’s GPT-4 is said to only have 1 trillion.
Mark Zuckerberg, CEO of Facebook, said that he expects future Llama versions to surpass proprietary competitors within the next year. He said that the Meta AI chatbot, powered by these models, was on course to become the most popular AI Assistant by the end this year. Already, hundreds of millions are using it.
The release coincides with tech companies racing to prove that their portfolios of resource-hungry, large language models are capable of delivering significant gains in areas such as advanced reasoning that can justify the massive sums invested in them.
Meta’s top AI scientist believes that such models will run into limits in reasoning, and other types of AI will be required to make breakthroughs.
The company announced that in addition to its flagship model with 405 billion parameters, Meta will also release updated versions of the lighter weight 8 billion and 70 billion parameter Llama 3 original introduced in the spring.
The three new models can be multilingual, and they are able to handle more complex user requests through an expanded “context” window. According to Ahmad Al-Dahle of Meta, the head of generative artificial intelligence, this will improve the experience for generating computer codes.
Al-Dahle, in an interview with Reuters, said that the larger context windows gave the models a memory similar to a long-term storage that helps them process multi-step requests.
Al-Dahle also said that his team was able to improve Llama 3’s performance in tasks like solving math problems, by using AI.
Meta offers its Llama model largely for free to developers. Zuckerberg believes this strategy will pay off with innovative products, a reduced dependence on potential competitors, and a greater level of engagement in the company’s social networks. However, some investors are concerned about the costs involved.
Meta will also benefit if developers choose to use their free models instead of paid ones. This would undermine the business models its competitors. Meta’s announcement touted improvements on math and knowledge tests, which may make this prospect more attractive.
Although it is difficult to measure progress in AI development, Meta’s test results appeared to indicate that its Llama 3 was almost equaling and, in some cases, bettering Anthropic’s Claude 3.5 Sonnet, and OpenAI GPT-4o which are generally regarded as two of the most powerful frontier models available.
Meta’s model scored 73.8 on the MATH benchmark for competition level math words problems. This compares to GPT-4o at 76.6, and Claude 3.5 Sonnet at 71.1.
The model scored 88.6 in MMLU. This benchmark covers dozens subjects including math, science, and humanities. GPT-4o also scored 88.7, and Claude 3.5 Sonnet also scored 88.3.
Meta researchers teased that they would be releasing “multimodal” models of Llama 3 later this year, which will layer video, image and speech capabilities over the core text model.
Early tests indicate that these models are “competitive” with multimodal models like Google’s Gemini 1,5 and Anthropic’s Claude 3.5 Sonnet.
