Instructions to use microsoft/Florence-2-base-ft with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use microsoft/Florence-2-base-ft with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="microsoft/Florence-2-base-ft", trust_remote_code=True)

# Load model directly
from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("microsoft/Florence-2-base-ft", trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained("microsoft/Florence-2-base-ft", trust_remote_code=True)

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use microsoft/Florence-2-base-ft with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "microsoft/Florence-2-base-ft"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "microsoft/Florence-2-base-ft",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/microsoft/Florence-2-base-ft

SGLang

How to use microsoft/Florence-2-base-ft with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "microsoft/Florence-2-base-ft" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "microsoft/Florence-2-base-ft",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "microsoft/Florence-2-base-ft" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "microsoft/Florence-2-base-ft",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use microsoft/Florence-2-base-ft with Docker Model Runner:
```
docker model run hf.co/microsoft/Florence-2-base-ft
```

Which one should I use, Florence-2-base-ft and Florence-2-base?

#23

by CrimisionLFIre - opened Dec 11, 2024

Discussion

CrimisionLFIre

Dec 11, 2024

In the description, I cant understand the main different between Florence-2-base-ft and Florence-2-base.
My question is, If I want to fine tuning a Florence-2 on specific object detection dataset, which one should I use, ft version or without ft version?

For now I use Florence-2-base-ft, but the results are not good enough because I found CNN based method can beat Florence-2-base-ft very easily, so I guess maybe there is something wrong during my fine-tuning phase.

defaultsettings

Feb 15, 2025

Hey I got he same question. Did you find the answer and what is the difference between the models?

ponkin

Jul 30, 2025

same question, what's the difference?

Novmik

Sep 29, 2025

•

edited Sep 29, 2025

Here is what GPT5-thinking summarized from the model description:

Short version: same architecture/size; different training stage.

Florence-2-base = the pre-trained checkpoint trained on Microsoft’s FLD-5B (5.4B annotations on 126M images). It’s the raw foundation model.
Florence-2-base-ft = the same model further fine-tuned on a curated mixture of downstream tasks (captioning, grounding/REC, detection, OCR, etc.) to make a single “generalist” model that performs better out-of-the-box on those tasks.

What that means in practice

Measured gains: Microsoft’s card shows the ft model improving typical benchmarks vs. the base model’s zero-shot numbers—for example, COCO Caption CIDEr 140.0 for base-ft (after fine-tune) vs. 133.0 for base (zero-shot); COCO detection mAP 41.4 (base-ft) vs. 34.7 (base zero-shot).
Under the hood: both are the same seq-to-seq VLM; the ft checkpoint is just an extra stage of supervised multi-task tuning layered on top of the FLD-5B pretraining described in the technical report.

Which to pick?

If you just want plug-and-play captions/OD/OCR, use Florence-2-base-ft. That’s what it’s for.
If you plan to fine-tune on your own data, starting from base is reasonable; some community users also report cases where base behaves better for niche prompts, but that’s anecdotal.

That’s the whole difference: pre-trained vs. the same model after multi-task fine-tuning—same size, same prompts, better out-of-box task performance on common datasets with the ft checkpoint.

If MS team has any comments feel free to pitch in

DasCaptain

Mar 2

thanks Novmik, you prompted chatGPT, i couldnt do that. Thanks for the knoweldge, you know a lot -.-

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment