Meta Platforms (META.O) has released the most advanced version of its Llama 3 artificial intelligence model, claiming significant improvements in multilingual capabilities, computer code generation, and complex math problem-solving. The announcement on Tuesday highlights Meta's push to rival paid models from competitors like OpenAI.

The new Llama 3 model, featuring 405 billion parameters, surpasses its predecessor while remaining smaller than leading models such as OpenAI's GPT-4, which has one trillion parameters, and Amazon's upcoming two trillion parameter model. Meta CEO Mark Zuckerberg anticipates that future Llama models will outperform proprietary competitors by next year, with the current Meta AI chatbot on track to become the most popular AI assistant by the end of this year.

In addition to the flagship model, Meta is releasing updated versions of its 8 billion and 70 billion parameter Llama 3 models. All versions are multilingual and can handle larger user requests through an expanded "context window," improving the generation of computer code.

Meta's head of generative AI, Ahmad Al-Dahle, highlighted that the larger context window addresses community feedback, providing the models with a longer memory to process multi-step requests. The Llama 3 model's performance on tasks like solving math problems has also improved, thanks to AI-generated training data.

Meta's strategy of offering its Llama models free-of-charge aims to foster innovation, reduce reliance on competitors, and increase engagement on its social networks. While some investors are concerned about the associated costs, Meta expects to benefit if developers choose its free models over paid alternatives, potentially disrupting rivals' business models.

Test results provided by Meta suggest that the largest Llama 3 model is nearly matching, and in some cases surpassing, Anthropic's Claude 3.5 Sonnet and OpenAI's GPT-4o in key benchmarks. On the MATH benchmark, Meta's model scored 73.8, compared to GPT-4o's 76.6 and Claude 3.5 Sonnet's 71.1. On the MMLU benchmark, Meta's model scored 88.6, close to GPT-4o's 88.7 and Claude 3.5 Sonnet's 88.3.

Meta researchers also hinted at upcoming multimodal versions of the Llama 3 models, incorporating image, video, and speech capabilities, expected to perform competitively with models like Google's Gemini 1.5 and Anthropic's Claude 3.5 Sonnet.