Meta's largest AI model .. free and multilingual

Meta Inc. on Tuesday released the largest version of its AI model (Lamma 3), which is mostly free and multilingual and has general performance metrics that compete with paid models from rival companies, such as OpenAI.

Meta, which owns Facebook, said in a blog post and a research paper announcing the release that the new Lama 3 model can speak eight languages, write high-quality computer code, and solve more complex mathematical problems than previous versions.

Last year's version pales in comparison to this version, which includes 405 billion metrics or variables that algorithms take into account to generate responses to user queries, although it is still smaller than the leading models from competitors.

There are reports that OpenAI's GPT-4 model includes a trillion variables, and Amazon is investing in a model with two trillion.

Meta Platforms said that in addition to its flagship Lama 3 model, it will release two updated versions of its shortened models, one with 8 billion variables and the other with 70 billion, which it introduced in the spring.

The three new models are multilingual and can handle more user requests via expanded “context windows.”

They will particularly improve the experience of generating computer code, said Ahmed Aldahleh, Meta’s head of generative AI.

“That was the first feedback we got from the community,” Aldahleh told Reuters in an interview, noting that expanded context windows give models something akin to a longer-term memory that helps them process multi-step requests.

Meta largely lets developers use Lama models for free, a strategy that CEO Mark Zuckerberg says will pay off in the form of innovative products and greater engagement on the company’s core social networks. However, the costs involved have raised eyebrows among some investors.

The company would also benefit if developers chose to use its free models instead of paid ones, which would limit the use of competitors’ models.

Meta touted gains in key math and cognitive tests in its announcement that could make this prospect more attractive.

While progress in AI development is difficult to measure, Meta’s test results appear to suggest that its largest model, LAMA, is nearly identical to Anthropic’s Cloud 3.5 Sonnet and OpenAI’s GPT-O4, and in some cases even outperforms them.

The release comes as tech companies race to show that their growing portfolios of large, resource-hungry language models can deliver big enough gains in problematic areas like advanced reasoning to justify the huge sums invested in them.