Gemma is a family of lightweight, state-of-the-art open models built from the same research and technology used to create the [Gemini](https://deepmind.google/models/gemini/) models. Developed by Google DeepMind and other teams across Google, Gemma is named after the Latin *gemma*, meaning precious stone. The Gemma model weights are supported by developer tools that promote innovation, collaboration, and the responsible use of artificial intelligence (AI).
The latest in the Gemma family is [Gemma 4](../models/gemma-4/index.md) featuring multimodal (text, image, audio) inputs.
!!! tip "New to Gemma?"
To learn more about the Gemma family of models, visit the [model family](../models/index.md) page for a tour.
## Choose a model
Gemma 4 ships in five sizes. They share a tokenizer and prompt format, so you can start small
and move up without rewriting anything.
| Size | Input ➔ Output | Context | Intended Platform |
|---|---|---|---|
| **E2B** | Text, Image, Audio, Video ➔ Text | 128K | Mobile devices |
| **E4B** | Text, Image, Audio, Video ➔ Text | 128K | Mobile devices and laptops |
| **12B** | Text, Image, Audio, Video ➔ Text | 256K | Laptops, desktop computers and small servers |
| **26B A4B** | Text, Image, Video ➔ Text | 256K | Desktop computers and small servers |
| **31B** | Text, Image, Video ➔ Text | 256K | Large servers or server clusters |
## Run your first inference
=== "Hugging Face Transformers"
Install [`transformers`](https://huggingface.co/docs/transformers/installation), then run:
```python { .model-swap }
from transformers import pipeline
pipe = pipeline(
"text-generation",
model="google/gemma-4-12B-it",
device_map="auto"
)
output = pipe("Explain quantum computing simply:", max_new_tokens=150)
print(output[0]["generated_text"])
```
=== "llama.cpp"
Install [`llama.cpp`](https://llama.app/), then run:
```bash { .model-swap }
llama-cli -hf google/gemma-4-12B-it-qat-q4_0-gguf --prompt "Write a poem about the Kraken."
```
=== "Ollama"
Install [`ollama`](https://ollama.com/download), then run:
```bash { .model-swap }
ollama run gemma4:12b "Explain quantum computing simply:"
```
=== "Apple MLX"
Install [`MLX LM`](https://github.com/ml-explore/mlx-lm), then run:
```bash { .model-swap }
mlx_lm.generate --model mlx-community/gemma-4-12B-it-qat-4bit \
--prompt "Explain quantum computing simply:"
```
=== "vLLM"
Install [`vLLM`](https://docs.vllm.ai/en/latest/getting_started/installation/), then run:
```bash { .model-swap }
vllm serve google/gemma-4-26B-A4B-it
```
=== "LMStudio"
Install [`LMStudio`](https://lmstudio.ai/), then run:
```bash { .model-swap }
lms chat google/gemma-4-12b
```
!!! example "Using a coding agent?"
Install the gemma-dev skill so your agent stays current with the latest models:
```bash
npx skills add google-gemma/gemma-skills --skill gemma-dev --global
```
## Providers
If you want to run Gemma 4 on a different stack, you can choose from the following:
## Get the weights
Every size ships full-precision weights plus [Quantization-Aware Training](qat.md) (QAT) checkpoints,
which hold up far better at 4-bit than a generic post-training quant.
| Where | What's there |
|---|---|
| [Hugging Face][hf-gemma4] | Safetensors, GGUF and QAT collections |
| [Kaggle][kaggle-gemma4] | The same checkpoints, mirrored |
| [Ollama][ollama-gemma4] | Prebuilt tags, pulled on first run |
Checkpoint names follow `google/gemma-4-{SIZE}-it{SUFFIX}` — pick the suffix your engine wants:
| Running on | Suffix |
|---|---|
| llama.cpp, LM Studio, Ollama | `-qat-q4_0-gguf` |
| vLLM, SGLang | `-qat-w4a16-ct` |
| Mobile | `-qat-mobile-transformers`, `-qat-mobile-ct` |
## What's next
--8<-- "glossary.md"