Source: Wall Street News
The report said that the official version of Llama 3 will support multi-modal processing, and the first two minor versions will not have this capability.
On Monday local time, the tech media The Information quoted one$Meta Platforms (META.US)$Employee news reports that the company plans to launch two small Llama 3 Major Language Models (LLMs) next week as a preliminary version of the official Llama 3 in the summer.
The release of these two small models is expected to pave the way for the official debut of Llama 3. Meta released Llama 2 in July of last year. Since then, several companies, including Google, Musk's XAI, and Mistral, have released open source big language models, and competition is fierce.
Llama 3 directly targets OpenAI's GPT-4, which has become a powerful multi-modal model that not only processes longer text, but also supports image input.
According to the report, the official version of Llama 3 will also support multi-modal processing, that is, simultaneously understanding and generating text and images; the first two minor versions are not capable of this.
Generally speaking, smaller models cost less and run faster, especially at a time when running large models is expensive, their value is particularly prominent. The small model also makes it easy for developers to develop artificial intelligence software on mobile devices.
Meta has previously launched three versions of Llama 2, the largest of which has 70 billion parameters, and the other two versions have 13 billion and 7 billion parameters, respectively.
According to a previous article on Hard AI, the largest version of Llama 3 could have more than 140 billion parameters.
Meta will also improve Llama 2's overly conservative response to controversial topics in Llama 3, and the researchers plan to loosen the big model's limitations in this regard so that it can interact more with users, provide context, and not just refuse to answer.
editor/tolk