Falcon 2: An 11B parameter VLM released with image understanding

AI Impact Summary

Falcon 2 is a new 11B parameter language model and VLM trained on a massive dataset of 5000B tokens across 11 languages. This release introduces a VLM variant capable of understanding and responding to queries about images, integrating a CLIP ViT-L/14 vision encoder. The model’s training strategy includes pretraining and finetuning stages, with a focus on dynamic encoding for detailed visual perception and instruction tuning on a large image-text dataset.