Meta is planning to unveil two smaller versions of its upcoming Llama 3 next week, reported The Information. These smaller models are expected to serve as a precursor to the launch of the largest version of Llama 3, anticipated this summer.

The release of these smaller models is aimed at generating excitement for the forthcoming Llama 3, which is scheduled to debut roughly a year after Llama 2 was launched last July. This move comes as several companies, including Google, Elon Musk’s xAI, Databricks and Mistral, have already introduced open-source LLMs. 

The Llama 3 project is part of Meta’s strategy to compete with OpenAI’s GPT-4, known for its ability to answer questions based on user-uploaded images. The upcoming biggest version of Llama 3 is expected to be multimodal, capable of processing both text and images. In contrast, the two smaller models set for release next week will lack multimodal capabilities, as per the Information report.

Smaller models are increasingly valued in the industry due to their cost-effectiveness and faster processing speeds compared to larger counterparts. They are particularly attractive for developers aiming to integrate artificial intelligence software into mobile devices.

Previously, Meta released three variants of Llama 2, ranging from 7 billion to 70 billion parameters, which encode the learning acquired during model training. 

It is speculated that the largest version of Llama 3 could exceed 140 billion parameters, as reported by The Information.

Meta utilises Llama 2 to power its AI assistant across its apps. Recent efforts within Meta’s generative AI department have focused on making Llama 3 more adept at addressing controversial queries, following concerns that Llama 2 was overly conservative in its responses.