Idefics2
the six Ws · specification
Hugging Face's M4 research team created Idefics2.
An 8-billion-parameter open multimodal model for image-text dialogue and document understanding.
Hosted on Hugging Face under the HuggingFaceM4 organization.
Idefics2 was released in April 2024, succeeding Idefics released in 2023.
It provided a fully open, reproducible alternative to Flamingo-style multimodal models with transparent training data.
Runs on transformers, combining a Mistral-7B language model with a SigLIP vision encoder.
Open-weight release by Hugging Face under Apache 2.0 license, with fully documented training data.
works with