A Quick Overview of Google’s Gemma Models

Gemma is a family of lightweight LLMs released by Google. It uses the same technology stack that Google used to build Gemini models.

On Kaggle, Gemma is easy to discover and try. In Kaggle’s model pages, Gemma appears near the top, and its growth trend is also strong. The model page is:

https://www.kaggle.com/models/google/gemma

At the time of the original note, Gemma had 640 code examples, 62 discussions, and 11 related competitions on Kaggle. Those examples and discussions are useful learning material if you want to understand how people run and fine-tune Gemma in practice.

The Gemma page also provides download information. There are versions for different frameworks, including Keras, PyTorch, Transformers, Gemma C++, and LiteRT. Each framework may contain several model variants. For example, the PyTorch release includes variants such as 2b, 2b-it, 7b, and 7b-it.

The 2b and 7b names refer to the rough parameter scale: 2 billion and 7 billion parameters. Each size has a base pretrained model and an instruction-tuned model, usually marked with the -it suffix.

The smallest 2b model can be tried even on a laptop without a dedicated GPU. Larger 7b models generally need more GPU memory. As a rough early reference, the 7B instruction-tuned model and full 7B model may require around 8GB to 16GB of VRAM depending on runtime and precision. Google’s later Gemma 2 family goes larger, up to 27B parameters, but it also needs stronger hardware.

Kaggle is a convenient first place to try Gemma. It also provides a limited amount of free GPU and TPU time each week. If you want to train or run Gemma locally, download the variant you need and follow the setup notes in the other articles in this series.

WeChat

Follow ZiCode on WeChat

If this post was useful, you can follow later updates on WeChat as well.

X / Twitter

Follow @ax2_zicode

Faster technical notes, short thoughts, and new-post alerts are posted on X.

X Open X Profile @ax2_zicode