Hugging Face releases IDEFICS, an open Flamingo reproduction

Hugging Face released IDEFICS, an open-access vision-language model that takes interleaved sequences of images and text and generates text in response — answering questions about an image, describing its contents, or producing a narrative grounded in several images. It came in two sizes, 9 billion and 80 billion parameters, each in a base and an instruction-tuned version.

IDEFICS was built as a reproduction of DeepMind’s Flamingo, a visual-language model DeepMind had described in a 2022 paper but never released publicly. Hugging Face’s team trained IDEFICS using only publicly available data and models — including a new dataset they built for the purpose, OBELICS, containing 141 million interleaved image-text documents drawn from the web — and reported that IDEFICS achieved performance comparable to Flamingo across several benchmarks despite the reproduction effort having no access to DeepMind’s original training data or code.

The release was explicitly framed around transparency: Hugging Face’s announcement argued that state-of-the-art model development should be more open, in contrast to the practice, increasingly common among frontier labs, of publishing a capability paper without releasing weights, code or training data. IDEFICS demonstrated that a well-resourced open effort could close much of that gap using only public resources, and it became a reference point and building block for subsequent open multimodal model work.