Introducing Idefics2: A Powerful 8B Vision-Language Model for the community

Hugging Face Blog · 2 years ago

Hugging Face released Idefics2, an 8-billion-parameter open-source vision-language model that processes text and images to answer questions, describe visual content, and extract information from documents. The model achieves performance on Visual Question Answering benchmarks comparable to much larger models like LLaVa-Next-34B and MM1-30B-chat, while being 10 times smaller than its predecessor Idefics1. Idefics2 is immediately available in the Hugging Face Transformers library with an Apache 2.0 license, enabling researchers to fine-tune it for specific multimodal applications.