> For the complete documentation index, see [llms.txt](https://docs.roboflow.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.roboflow.com/models/supported-models/paligemma2.md).

# PaliGemma 2

PaliGemma 2 is Google's vision-language model. It accepts an image and a text prompt and returns a text response. We support PaliGemma 2 through our [Serverless Cloud API](https://docs.roboflow.com/deployment/roboflow-cloud/serverless-api), [Dedicated Deployments](https://docs.roboflow.com/deployment/roboflow-cloud/dedicated-deployments), and [self-hosted Inference](https://docs.roboflow.com/deployment/self-hosted/self-hosted).

## PaliGemma 2 API

{% stepper %}
{% step %}

### Get your API Key

Create a Roboflow account, find your key on the [Roboflow API settings page](https://app.roboflow.com/settings/api) and make it available to your shell:

```bash
export ROBOFLOW_API_KEY="your-key-here"
```

{% endstep %}

{% step %}

### Install the dependencies

Install the [Inference SDK](https://docs.roboflow.com/deployment/self-hosted/self-hosted):

```bash
pip install -U inference-sdk supervision
```

{% endstep %}

{% step %}

### Run the model

The sample calls the pretrained `paligemma2-3b-pt-224` checkpoint with a caption prompt.

```python
import os
import supervision as sv
from inference_sdk import InferenceHTTPClient

image = sv.load_image_from_url("https://media.roboflow.com/quickstart/dog.jpeg")

client = InferenceHTTPClient(
    api_url="https://serverless.roboflow.com",
    api_key=os.environ["ROBOFLOW_API_KEY"],
)
result = client.infer_lmm(
    image,
    model_id="paligemma2-3b-pt-224",
    prompt="caption en",
    max_new_tokens=64,
)
print(result["response"])
```

{% endstep %}
{% endstepper %}

The code above prints the model response to the terminal:

```
a dog is seen here on the shoulder of a man
```

<figure><img src="/files/yBX3kRjJOf7fkJHCGLuT" alt=""><figcaption></figcaption></figure>

## PaliGemma 2 inference speed

Latency measured with [Roboflow Inference](https://docs.roboflow.com/deployment/self-hosted/self-hosted) on 1x NVIDIA L4, batch size 1, generating exactly 128 tokens with greedy decoding from a fixed prompt. Latency scales with output length, so use tokens/sec to estimate other lengths.

<table data-search="false"><thead><tr><th>Alias</th><th>Latency, 128 tokens (ms)</th><th>Tokens/sec</th></tr></thead><tbody><tr><td><code>paligemma2-3b-pt-224</code></td><td>3986</td><td>32</td></tr></tbody></table>

{% hint style="info" %}
Set `api_url` to match your deployment target:

* `https://serverless.roboflow.com` for the Serverless Cloud API.
* `http://localhost:9001` for a local [Inference](https://docs.roboflow.com/deployment/self-hosted/self-hosted) server.
* Your [Dedicated Deployment](https://docs.roboflow.com/deployment/roboflow-cloud/dedicated-deployments) URL for a private endpoint.
  {% endhint %}

You can train your own PaliGemma 2 checkpoint on Roboflow and call it by its per-model `{workspace}/{model-slug}` ID (see [Versions, Trainings, and Models](/models/versions-trainings-and-models.md)). See the [Inference documentation](https://docs.roboflow.com/deployment/self-hosted/self-hosted) for additional prompt formats and supported checkpoints.

## PaliGemma 1 (legacy)

The original PaliGemma release is still loadable through the [`inference`](https://docs.roboflow.com/deployment/self-hosted/self-hosted) package on your own hardware. New projects should use PaliGemma 2 above; this section is kept for existing integrations.

Install the package:

```bash
pip install "inference[transformers]"
```

Use `inference-gpu[transformers]` on a GPU machine.

### Visual question answering

```python
from PIL import Image

from inference.models.paligemma.paligemma import PaliGemma

model = PaliGemma("paligemma-3b-mix-224", api_key="YOUR_API_KEY")

image = Image.open("image.jpeg")
result = model.predict(image, "How many dogs are in this image?")

print(result)
```

### Object detection

PaliGemma emits detections as `<loc####>` tokens rather than JSON, so the response has to be parsed before it can be visualized. Prompt with `detect <class>; <class>` and decode the tokens into boxes:

```python
import re
from typing import List, Optional, Tuple

import numpy as np
import supervision as sv

_DETECT_RE = re.compile(r"(.*?)" + r"<loc(\d{4})>" * 4 + r"\s*([^;<>]+)? ?(?:; )?")


def from_pali_gemma(
    response: str,
    resolution_wh: Tuple[int, int],
    class_list: Optional[List[str]] = None,
) -> sv.Detections:
    width, height = resolution_wh
    xyxy_list, class_name_list = [], []

    while response:
        match = _DETECT_RE.match(response)
        if not match:
            break

        groups = list(match.groups())
        before = groups.pop(0)
        name = groups.pop()
        y1, x1, y2, x2 = [int(value) / 1024 for value in groups[:4]]
        y1, x1, y2, x2 = map(round, (y1 * height, x1 * width, y2 * height, x2 * width))

        content = match.group()
        if before:
            response = response[len(before):]
            content = content[len(before):]

        xyxy_list.append([x1, y1, x2, y2])
        class_name_list.append(name.strip())
        response = response[len(content):]

    class_name = np.array(class_name_list)
    class_id = (
        np.array([class_list.index(name) for name in class_name])
        if class_list is not None
        else None
    )
    return sv.Detections(
        xyxy=np.array(xyxy_list),
        class_id=class_id,
        data={"class_name": class_name},
    )


classes = ["person", "car", "backpack"]
response = model.predict(image, "detect person; car; backpack")[0]
detections = from_pali_gemma(response, resolution_wh=image.size, class_list=classes)
```

Pass the resulting `sv.Detections` to [supervision](https://supervision.roboflow.com/) annotators to draw the boxes.