---
pipeline_tag: text-generation
license: gemma
base_model: google/gemma-2b-it
library_name: zeromodels
extra_gated_heading: Access Gemma on Hugging Face
language:
- en
tags:
- keras
- zeromodels
- gemma
- gemma-2b
- text-generation
- arxiv:2403.08295
- pytorch
- jax
- tf
---

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

# Run Gemma 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-1f6feb)](https://imvision12.github.io/ZeroModels/gemma/) [![HuggingFace](https://img.shields.io/badge/HuggingFace-Gemma-ffd21e?logo=huggingface&logoColor=black)](https://huggingface.co/zeromodels)

# zeromodels/gemma-2b-it

Pure-**Keras 3** conversion of [`google/gemma-2b-it`](https://huggingface.co/google/gemma-2b-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 **text -> text** via `GemmaTextGenerate`; weights are
stored in **bfloat16**.

For model details, license, and usage terms, see Google's
[model card](https://huggingface.co/google/gemma-2b-it).

## ✨ Quick start

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

from zeromodels.models.gemma import GemmaTextGenerate, GemmaTokenizer

model = GemmaTextGenerate.from_weights("zeromodels/gemma-2b-it")
tokenizer = GemmaTokenizer.from_weights("zeromodels/gemma-2b-it")

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

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

| Variant | Hub |
| --- | --- |
| `gemma-1.1-2b-it` | [zeromodels/gemma-1.1-2b-it](https://huggingface.co/zeromodels/gemma-1.1-2b-it) |
| `gemma-1.1-7b-it` | [zeromodels/gemma-1.1-7b-it](https://huggingface.co/zeromodels/gemma-1.1-7b-it) |
| `gemma-2b` | [zeromodels/gemma-2b](https://huggingface.co/zeromodels/gemma-2b) |
| `gemma-2b-it` | [zeromodels/gemma-2b-it](https://huggingface.co/zeromodels/gemma-2b-it) |
| `gemma-7b` | [zeromodels/gemma-7b](https://huggingface.co/zeromodels/gemma-7b) |
| `gemma-7b-it` | [zeromodels/gemma-7b-it](https://huggingface.co/zeromodels/gemma-7b-it) |

## 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 docs](https://imvision12.github.io/ZeroModels/gemma/).
- Community / upstream weights still work via the `hf:` prefix:
  `GemmaTextGenerate.from_weights("hf:google/gemma-2b-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.