Just two months after the release of the large language model Llama 3.1, Meta
introduced its updated version Llama 3.2 – the first open artificial intelligence system that can process both images and text.
Image source: Gerd Altmann / pixabay.com
Meta Llama 3.2 allows developers to create advanced AI applications: augmented reality platforms with real-time video recognition; visual search engines with content-based image sorting; as well as document analysis systems with the preparation of summaries of long text fragments. It will be quite simple for developers to launch a new model, they say in Meta – they will need to add support for multimodality, “be able to show Llama images and make it communicate.”
OpenAI and Google launched their own multimodal AI models last year, so Meta now finds itself in a catching-up position. Imaging support is important as Meta continues to expand AI capabilities across devices, including Ray-Ban Meta glasses. The Llama 3.2 package includes two models that support working with images (with 11 and 90 billion parameters) and two lightweight text models (with 1 and 3 billion parameters). The smaller ones are designed to run on chips from Qualcomm, MediaTek and other Arm processors – Meta clearly expects them to be used on mobile devices. That said, Llama 3.1, released in July, is still a strong offering – one version has 405 billion parameters, and should outperform newer ones in text generation.