Idefics3 · Hugging Face · Model History

Claims and evidence

Released August 2024No dated announcement. The month follows from the Hub repository (created 2024-08-05; first public community discussion 2024-08-08) and the technical report (arXiv v1, 2024-08-22). Primary source[1]Hub commit history: repository created 2024-08-05; community tab: first discussion 2024-08-08 [2]arXiv v1 of 22 Aug 2024; Introduction: model released alongside the training datasets

Successor of Idefics2 Primary source[1]introduction: improves upon Idefics1 and Idefics2 [2]Introduction: its predecessor, Idefics2-8B
Derived from (other) Llama 3.1 (Meta) Primary source[1]parent model: Llama 3.1 8B Instruct

Change · Architecture Uses Llama 3.1 Instruct instead of Mistral-7B as the language model.Compared with Idefics2 Primary source[2]
Change · Architecture Encodes each image tile of up to 364x364 pixels with 169 visual tokens; larger images are split into tiles that are encoded separately.Compared with Idefics2 Primary source[1]
Change · Training data Fine-tuning data extends The Cauldron with further datasets, including Docmatix.Compared with Idefics2 Primary source[1]
Change · Other Hugging Face reports a large improvement over Idefics2 in document understanding.Compared with Idefics2 Primary source[1]

Input text, image Primary source[1]introduction; Model Summary: Model type
Output text Primary source[1]introduction
Open weights Yes Primary source[1]introduction; License: checkpoints released under Apache 2.0
Parameters 8B Primary source[2]Introduction: Idefics3-8B [1]Uses: Idefics3-8B
Access Open-weights download Primary source[1]How to Get Started

No known successor.

None recorded.

None recorded.

None.

AI Radar coverage starts in June 2026; no coverage linked yet.