Meta may launch a compact Llama AI model ahead of the larger version

These models will be released prior to the launch of the flagship model later this year.

Allegedly, Meta plans to introduce scaled-down versions of its Llama language model, catering to the growing demand for more economical AI models among businesses.

The Information reports that the company plans to launch two small Llama 3 versions this month before putting out the flagship model this summer. The Verge reached out to Meta for comment.

The move underscores the growing trend of AI developers adding lightweight AI model options. Meta already has a smaller version of its Llama 2 model, Llama 2 7B, which it launched in February last year. Google came out with the Gemma family of models in February, and the French AI company Mistral also has Mistral 7B.

These models typically cannot handle long strings of instructions from users but are faster, more flexible, and, most importantly, cheaper to run than a regular-sized model. But these are still powerful AI models, able to summarize PDFs and conversations and write code. Larger models are usually used for more complicated tasks like generating photos or tasks that require several commands to execute. Since small models only work with a smaller number of parameters (data that it learns), these also require less computing power and, therefore, are more cost-effective.

Lightweight models tend to attract users who don’t necessarily want to use the breadth of a large language model for their applications. Smaller models can most often be deployed in specific projects like code assistance or in devices that cannot handle the power usage of a bigger AI model, like phones or laptops.

Meta reportedly plans a July release for Llama 3, which may be “looser” than the previous version and be able to answer controversial questions Llama 2 was not allowed to answer.

Source : The Verge