Meta Unveils Llama 3.2 Multimodal and On-Device Models

Meta has released Llama 3.2, which includes multimodal capabilities and on-device optimized models. This update includes a total of 10 open-weight models.

Llama 3.2 Vision is available in two sizes. There is an 11B model that can run efficiently on consumer-grade GPUs, and a 90B model for large-scale applications, both showing excellent performance in tasks such as visual reasoning, document Q&A, and image-text retrieval.

Additionally, small text-only models in 1B and 3B sizes have been added, capable of running on-device. For safety, Llama Guard 3, which can detect multimodal prompts and responses, has also been released.

Hugging Face supports integration with Transformers and TGI, as well as deployment and fine-tuning through various infrastructures such as Google Cloud, Amazon SageMaker, and DELL Enterprise Hub.