Meta Platforms on Tuesday released the biggest version of its mostly free Llama 3 artificial intelligence model, boasting multilingual skills and general performance metrics that rival paid models from rivals such as OpenAI.
The new Llama 3 model can converse in eight languages, write high-quality computer code and solve more complex math problems than previous versions, Facebook’s parent company said in a blog post and a research paper announcing its release.
With 405 billion parameters, or variables, that the algorithm takes into account to formulate answers to user queries, it is smaller than the previous version released last year, though it is still smaller than the leading models offered by competitors.
In contrast, OpenAI’s GPT-4 model is reported to have one trillion parameters, and Amazon is building a model with 2 trillion parameters.
Promoting Llama 3 on multiple channels, Chief Executive Mark Zuckerberg said he expects future Llama models to outperform proprietary competitors by next year. He said the Meta AI chatbot powered by those models is on track to become the most popular AI assistant by the end of this year, already used by hundreds of millions of people.
The release comes at a time when technology companies are trying to show that their growing portfolio of resource-hungry large language models can deliver substantial gains in known problem areas, such as advanced reasoning, to justify the large sums invested in them.
Meta’s own top AI scientist has said he believes such models will hit the limits of logic and will require other types of AI systems to achieve success.
In addition to its flagship 405 billion parameter model, Meta is also releasing updated versions of its lighter-weight 8 billion and 70 billion parameter Llama 3 models, which were initially introduced in the spring, the company said.
All three new models are multilingual and can handle larger user requests through an expanded “context window,” which Ahmed Al-Dahle, Meta’s head of generative AI, said will particularly improve the experience of creating computer code.
“This was the best feedback we received from the community,” Al-Dahle told Reuters in an interview, adding that the larger reference window gives models a longer memory-like feature, which is helpful when processing multi-step requests.
In addition, Al-Dahle said his team has succeeded in improving the performance of the Llama 3 models on tasks such as solving math problems, by using AI to generate some of the data on which they were trained.
Meta releases its Llama model largely free of charge for use by developers, a strategy that Zuckerberg says will yield benefits in the form of innovative products, less reliance on potential competitors and more engagement on the company’s main social network. However, some investors have raised their eyebrows over the cost.
The company will also benefit if developers use its free model rather than a paid one, which would undermine its competitors’ business models. With its announcement, Meta claimed advantages on major math and knowledge tests that could make that prospect even more attractive.
Although the progress of AI development is notoriously difficult to measure, test results provided by Meta suggest that its largest Llama 3 model is nearly matching and in some cases outpacing Anthropic’s Cloud 3.5 Sonnet and OpenAI’s GPT-4o, which are widely regarded as the two most powerful frontier models on the market.
For example, on the MATH benchmark of competition-level math word problems, Meta’s model posted a score of 73.8, while GPT-4o’s was 76.6 and Cloud 3.5 Sonnet’s was 71.1.
The model scored 88.6 on MMLU, a benchmark covering dozens of topics across math, science, and humanities, while GPT-4o scored 88.7 and Cloud 3.5 Sonnet scored 88.3.
In their paper, the Meta researchers also hint at “multimodal” versions of the model coming later this year, which will incorporate image, video, and speech capabilities on top of the core Llama 3 text model.
He said initial experiments show that these models can perform “competitively” with other multimodal models such as Google’s Gemini 1.5 and Anthropic’s Cloud 3.5 Sonnet.
(Except for the headline, this story has not been edited by NDTV staff and is published from a syndicated feed.)
