Meta has released its new one large language model (LLM) that is to say Llama 3.
This, in addition to being free for common users, it will also be available without any cost for thecommercial use for contexts with less than 700 million users active per month. In reality we are talking about two models, both pre-trained and capable of managing the most commonly related tasks.Artificial intelligencesuch as the management of chatbot.
According to the data provided by Meta, Llama 3 should show performances that, in different contexts, prove to be clearly superior to the competition, i.e. Mistral, Gemini e Claude 3. The IT giant has worked on improving the post-training procedure, offering greater diversity of responses.
The new LLM is designed to deliver optimized performancecon much more accurate answers compared to what was proposed by Llama 2. Compared to its predecessor, however, there are also other substantial differences.
Llama 3 was powered with seven times the data used for Llama 2
Llama 3, in fact, is able to work with a greater number of tokenwith a limit carried from meta to 128.000. In fact, the new system is able to encode language more efficiently.
Still in the context of tokens, more than were used to train Meta’s new AI 150 trillion tokens, collected from public sources. To better quantify this amount of data, just think that it is the same seven times larger than Llama 2.
Not only that: to facilitate work on languages in addition toEnglishMeta developers used the 5% of data coming from multilingual sources (30 different languages). This means that while Llama 3 will perform best in English, it will be able to interact with other languages more effectively than before. According to the company, Llama models are a part of a larger ecosystem in which developers take a privileged role.
In all likelihood, Meta’s run is not destined to end here: new models will be offered in the future, with even more surprising features. The company, in this sense, spoke of greater flexibility regarding languages other than English, larger context windows and, in general, more efficient LLMs.
We just have to wait a few months to see what Meta has in store to surprise us again.