Florence-2 Large

Florence-2 Large is the 770-million-parameter variant of Microsoft Research's Florence-2 vision foundation model, released June 2024. It treats diverse vision tasks as a single sequence-to-sequence problem: the same network produces captions, bounding boxes, segmentation masks, or OCR output depending on the text prompt it receives.

The model was trained on FLD-5B, Microsoft's internal dataset of 126 million images with 5.4 billion annotations, which lets this sub-billion-parameter model match or beat much larger specialists on COCO captioning, referring expression comprehension, and detection benchmarks. A 230M base sibling ships for even tighter footprints.

Weights are released under the permissive MIT license on Hugging Face, with no commercial-use restrictions, and the model is also served through Azure AI Foundry. Its unified prompt interface and tiny size make it a popular backbone for vision features that do not need a full multimodal LLM.