Idefics3-8B

Idefics3-8B is an open vision-language model released by Hugging Face researchers on August 22, 2024. It adapts the Idefics2 design with three key changes: a Llama 3 text backbone, updated image processing logic, and removal of the perceiver module. The model accepts arbitrary sequences of image and text inputs and produces text outputs.

Training ran exclusively on open datasets, including Docmatix, a new document-understanding dataset 240 times larger than previously available resources. It scores 87.7% on DocVQA and 55.9 on MMStar in vendor-reported evaluations, a large jump over Idefics2. Checkpoints ship under Apache 2.0 and the model is integrated into Hugging Face Transformers.