README.md

---
license: mit
library_name: coreml
pipeline_tag: image-text-to-text
base_model: microsoft/Florence-2-base
base_model_relation: quantized
tags:
  - coreml
  - core-ml
  - ios
  - macos
  - apple
  - on-device
  - vision-language
  - captioning
  - ocr
  - grounding
  - arxiv:2311.06242
---

# Florence-2 — Core ML

*Microsoft, 2024*

Vision-language captioning, OCR, and VQA. Three-stage encoder-decoder. 768×768 input.

<p><img src="https://huggingface.co/mlboydaisuke/Florence-2-base-CoreML/resolve/main/media/983d3432cc.png" alt="Florence-2 demo" width="49%"> <img src="https://huggingface.co/mlboydaisuke/Florence-2-base-CoreML/resolve/main/media/6e404a09a4.jpg" alt="Florence-2 demo" width="49%"></p>

Core ML conversion of [microsoft/Florence-2](https://huggingface.co/microsoft/Florence-2-base) for on-device inference on iPhone, iPad and Mac. Converted with `coremltools`; the packages are stateless, so all sequencing and buffering lives in your Swift code.

| | |
|---|---|
| Task | image text to text |
| Upstream | [microsoft/Florence-2](https://huggingface.co/microsoft/Florence-2-base) |
| Packages | 3 |
| Download size | 229 MB |
| Minimum iOS | 17.0 |
| Peak RAM | ~1200 MB |

## Files

| File | Size | Compute units | SHA-256 |
|---|---:|---|---|
| `Florence2VisionEncoder.mlpackage.zip` | 77 MB | `cpuOnly` | `9422f189c21220a0…` |
| `Florence2TextEncoder.mlpackage.zip` | 69 MB | `cpuOnly` | `f985deeef0408ea8…` |
| `Florence2Decoder.mlpackage.zip` | 81 MB | `cpuOnly` | `fe85a6faab528127…` |
| `florence2_vocab.json` | 976 KB | `-` | `861fee9af5520403…` |
| **Total** | **229 MB** | | |

`compute_units` is not a suggestion -- it is the configuration the conversion was verified against. Moving a package to a different compute unit can silently change the numerics (FP16 attention overflow) or crash on the GPU.

## Download

```bash
hf download mlboydaisuke/coreml-zoo --include "florence2/*" --local-dir ./florence2
unzip './florence2/florence2/*.zip' -d ./florence2
```

## Use in Swift

```swift
import CoreML

let config = MLModelConfiguration()
config.computeUnits = .cpuOnly   // as converted — see the table above

// Unzip the .mlpackage, drop it into your Xcode target and Xcode compiles it
// at build time:
let model = try Florence2VisionEncoder(configuration: config)

// ...or compile a downloaded .mlpackage at runtime:
let compiled = try await MLModel.compileModel(at: mlpackageURL)
let model = try MLModel(contentsOf: compiled, configuration: config)
```

> This model is split into 3 Core ML packages that are driven in sequence from Swift. Load them one at a time, copy the outputs out of the `MLMultiArray` buffers and release each model before loading the next — two large Core ML models resident at once will OOM on an iPhone.

## Demo

- **Sample app** — [`sample_apps/Florence2Demo`](https://github.com/john-rocky/CoreML-Models/tree/master/sample_apps/Florence2Demo), a standalone SwiftUI project.
- **Models Zoo** — this model is downloadable and runnable inside the [Models Zoo app](https://apps.apple.com/app/id6762083207) on the App Store, no build required.

## Conversion

- Script: [`convert_florence2.py`](https://github.com/john-rocky/CoreML-Models/blob/master/conversion_scripts/convert_florence2.py)
- Pitfalls hit during conversion (FP16 overflow, ANE buffer limits, stride handling): [`docs/coreml_conversion_notes.md`](https://github.com/john-rocky/CoreML-Models/blob/master/docs/coreml_conversion_notes.md)
- Model index: [CoreML-Models](https://github.com/john-rocky/CoreML-Models)

## License

The conversion inherits the upstream license: **MIT**.

## Credits

- Upstream authors: [microsoft/Florence-2](https://huggingface.co/microsoft/Florence-2-base), 2024
- Core ML conversion: john-rocky (Daisuke Majima)