Meta reveals its largest AI model yet

The new (Llama 3) model can speak eight languages, write high-quality computer code and solve more complex math problems than previous versions, Facebook owner Meta said in a blog post and research paper announcing the release.

Last year’s version pales in comparison to this year’s, which has 405 billion metrics, or variables, that the algorithm considers to generate responses to user queries, though it’s still smaller than the leading model from competitors.

It is reported that OpenAI’s GPT-4 model contains one trillion variables, and Amazon is investing in a model containing two trillion variables.

Tense competition

The launch comes as tech companies race to prove that their growing portfolios of large, resource-hungry language models can produce gains big enough in problem areas like high-level reasoning to justify the huge amounts of money invested in them.

Meta Platforms said that in addition to the flagship model, Llama 3, it will also launch two updated versions of its short models, one with 8 billion variables and the second with 70 billion variables, which it launched in the spring.

Meta has largely allowed developers to use the Llama model for free, a strategy that Chief Executive Mark Zuckerberg has said will pay off in the form of innovative products and increased engagement on the company’s core social media network. However, the costs involved have raised some concerns among investors.

The company also benefits if developers choose to use its free templates instead of paid ones, which will limit the use of competitor templates.

In her ad, Mehta touted improvements on key math and knowledge tests that might make the prospect more appealing.

Although it’s difficult to measure progress in AI development, the test results given by Meta seem to show that its largest model (LAMA) is almost the same as Anthropco’s Claude 3.5 Sonet model and OpenAI’s GPT-O4 model, or even outperforms them in some cases.