Meta has added two new ones linguistic models to his series: Llama 3 8B e Llama 3 70B. These promise superior performance in the world of artificial intelligence. The most striking news? The presence of 8 billion and 70 billion parameters respectively. These numbers represent the complexity and ability of the model to handle and to process information.
Llama 3 isn't just two models: more material is on the way
Meta doesn't stop there; she plans to release more models in the Llama 3 series, but without specifying a date. This strategy creates anticipation and also demonstrates the company's commitment to continually improving its offerings.
I new Llama models they are described as a leap from previous models. This improvement is due in part to the use of two custom GPU clusters, each with 24.000 units, which allowed models to be trained more effectively and accurately. Meta claims that, due to the number of parameters they possess, these models are among the best available today for generative artificial intelligence.
Llama 3's performance was measured through various AI benchmarks, which are standardized tests used to evaluate the capabilities of AI models. These include MMLU for knowledge, ARC for the acquisition of skills, e DROP for reasoning about blocks of text. While there are differing opinions on the usefulness of these benchmarks, they remain a common method for evaluating and comparing the capabilities of AI models.
In particular, Llama 3 8B outperformed similar open source models with fewer parameters, such as Mistral and Google Gemma, in at least nine different benchmarks. This includes tests on general knowledge, reasoning, and specific skills in areas such as biology, physics, chemistry, and even code generation and math problems. Which GPT-4, for example, cannot do very well.
The largest model, Llama 3 70B, not only competes but in some cases surpasses flagship models such as Gemini 1.5 Pro from Google in various tests, demonstrating that despite strong competition, it can stand up to even the giants of the sector. However, it is important to note that while it outperforms some flagship models such as Claude 3 Sonnet by Anthropic in several tests, it does not reach the performance of the more advanced Claude 3 Opus model.
