---
pipeline_tag: image-text-to-text
license: gemma
base_model: google/gemma-3-4b-it
library_name: zeromodels
extra_gated_heading: Access Gemma on Hugging Face
language:
- en
tags:
- keras
- zeromodels
- gemma3
- gemma-3
- image-text-to-text
- arxiv:2503.19786
- pytorch
- jax
- tf
---

*See [our collection](https://huggingface.co/zeromodels) for all Gemma 3 sizes and variants.*

# Run Gemma 3 with Keras 3: JAX, PyTorch, or TensorFlow

[![GitHub](https://img.shields.io/badge/GitHub-ZeroModels-181717?logo=github)](https://github.com/IMvision12/ZeroModels) [![Docs](https://img.shields.io/badge/Docs-Gemma_3-1f6feb)](https://imvision12.github.io/ZeroModels/gemma3/) [![HuggingFace](https://img.shields.io/badge/HuggingFace-Gemma_3-ffd21e?logo=huggingface&logoColor=black)](https://huggingface.co/zeromodels)

# zeromodels/gemma-3-4b-it

Pure-**Keras 3** conversion of [`google/gemma-3-4b-it`](https://huggingface.co/google/gemma-3-4b-it) for
[zeromodels](https://github.com/IMvision12/ZeroModels). One implementation runs unmodified on
**TensorFlow / Torch / JAX**. This is the instruction-tuned checkpoint, served here as **image + text -> text** via `Gemma3ConditionalGenerate`; weights are
stored in **bfloat16**.

For model details, license, and usage terms, see Google's
[model card](https://huggingface.co/google/gemma-3-4b-it).

## ✨ Quick start

### Text-only

```python
import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from zeromodels.models.gemma3 import Gemma3TextGenerate, Gemma3Tokenizer

model = Gemma3TextGenerate.from_weights("zeromodels/gemma-3-4b-it")
tokenizer = Gemma3Tokenizer.from_weights("zeromodels/gemma-3-4b-it")

inputs = tokenizer([{"role": "user", "content": "Hello, who are you?"}])
outputs = model.generate(**inputs, max_new_tokens=64)
print(tokenizer.decode(outputs[0]))
```

### Image + text

```python
from zeromodels.models.gemma3 import Gemma3ConditionalGenerate, Gemma3Processor

model = Gemma3ConditionalGenerate.from_weights("zeromodels/gemma-3-4b-it")
processor = Gemma3Processor.from_weights("zeromodels/gemma-3-4b-it")

conversation = [
    {"role": "user", "content": [
        {"type": "image", "url": "https://.../image.jpg"},
        {"type": "text", "text": "Describe this image."},
    ]},
]
inputs = processor(conversation)
outputs = model.generate(**inputs, max_new_tokens=64)
print(processor.decode(outputs[0]))
```

Load any Gemma 3 variant the same way with `from_weights("zeromodels/<variant>")`:

| Variant | Hub |
| --- | --- |
| `gemma-3-12b-it` | [zeromodels/gemma-3-12b-it](https://huggingface.co/zeromodels/gemma-3-12b-it) |
| `gemma-3-12b-pt` | [zeromodels/gemma-3-12b-pt](https://huggingface.co/zeromodels/gemma-3-12b-pt) |
| `gemma-3-1b-it` | [zeromodels/gemma-3-1b-it](https://huggingface.co/zeromodels/gemma-3-1b-it) |
| `gemma-3-1b-pt` | [zeromodels/gemma-3-1b-pt](https://huggingface.co/zeromodels/gemma-3-1b-pt) |
| `gemma-3-270m` | [zeromodels/gemma-3-270m](https://huggingface.co/zeromodels/gemma-3-270m) |
| `gemma-3-270m-it` | [zeromodels/gemma-3-270m-it](https://huggingface.co/zeromodels/gemma-3-270m-it) |
| `gemma-3-27b-it` | [zeromodels/gemma-3-27b-it](https://huggingface.co/zeromodels/gemma-3-27b-it) |
| `gemma-3-27b-pt` | [zeromodels/gemma-3-27b-pt](https://huggingface.co/zeromodels/gemma-3-27b-pt) |
| `gemma-3-4b-it` | [zeromodels/gemma-3-4b-it](https://huggingface.co/zeromodels/gemma-3-4b-it) |
| `gemma-3-4b-pt` | [zeromodels/gemma-3-4b-pt](https://huggingface.co/zeromodels/gemma-3-4b-pt) |

## Tips

- Set `KERAS_BACKEND` **before** importing Keras / zeromodels.
- Loads in **bfloat16** by default. Pass `load_dtype="float32"` for full precision,
  or `quantization="int8"` to shrink further.
- See the [Gemma 3 docs](https://imvision12.github.io/ZeroModels/gemma3/).
- Community / upstream weights still work via the `hf:` prefix:
  `Gemma3ConditionalGenerate.from_weights("hf:google/gemma-3-4b-it")`.

## Special Thanks

A huge thank you to the Google Gemma authors for creating and releasing these models.

License: Gemma (gated). Accept the license on the upstream Hub card before downloading.