Chief AI scientist Yann LeCun and others participated in the ‘Meta AI Day’ held in London. (Photo = Techcruch, Ingrid Lunden)
Meta’s large open source language model (LLM) ‘Rama 3’ will be released in three versions. Among them, two smaller versions are expected to be released next week.
The Information reported on the 8th (local time), citing an unnamed Meta employee, that two smaller versions of Rama 3 will be released next week.
According to this, the smallest model is said to be on-device AI for mobile devices. This model will be installed on Meta’s ‘smart Ray-Ban glasses’ and will support AI assistant functions.
In addition, the medium-sized model focuses on low operating costs and fast operating speed in line with recent trends. The parameters of the two models are not specifically known.
However, the largest version to be released this summer is said to have more than 140 billion parameters (140B). This is up to twice the size of ‘Rama 2’.
‘Rama 2’, which was launched in July last year, was also released in three versions, and the parameters were divided into 70 billion (70B), 13 billion (13B), and 7 billion (7B), respectively.
In particular, Rama 3 140B is known to have focused on catching up with ‘GPT-4’ performance. Therefore, it is said to be a multimodal model (LMM) that understands images beyond text. On the other hand, it was confirmed that the two small models were not multimodal.
Additionally, Meta reportedly deemed the current Rama 2 too conservative, and adapted Rama 3 to be able to answer controversial questions more openly. This is because Rama 3 works on meta AI chatbots that are served through Facebook, Instagram, etc.
Meanwhile, Meta’s Chief Product Officer, Chris Cox, also confirmed that the company plans to launch Rama 3 at an event to be held in London.
“We hope to launch a new family of Rama 3 within the next month, actually sooner, hopefully in a very short period of time,” he said. “Going forward this year, we will have a variety of models with different features coming out, and that will start very soon.”
However, in line with the trend of most companies being reluctant to reveal imaging functions after the ‘Gemini’ event, Meta was also cautious about including imaging functions.
Director Cox said, “When creating images, adequate testing for integrity and security is very important,” and said that the ‘Emu’ image creation tool was not included.
Correspondent Lim Da-jun ydj@aitimes.com
#Meta #release #small #version #Rama #week.. #big #version #billion #parameters
