Gemma

Overview

Gemma is an open model family launched by Google DeepMind, built on core technology from the same lineage as Gemini, for developers and researchers to create AI applications that can be deployed locally, fine-tuned, and commercialized. Gemma emphasizes being "lightweight + high performance" and can run on devices such as workstations, laptops, and even mobile phones, making it suitable for AI development, inference deployment, model research, and edge-side applications.

At present, the Gemma family shown on the official website is no longer limited to the early 2B and 7B models, but has also expanded into directions such as Gemma 3 and Gemma 3n, and includes derivative models for tasks such as translation, function calling, embeddings, and healthcare.

Main Features

• Open model weights: Provides open weights for convenient research, fine-tuning, and deployment, suitable for building proprietary AI capabilities.
• Lightweight inference: The models are optimized for a variety of hardware environments and can run flexibly on local devices and in the cloud.
• Multiple model versions available: Offers pretrained and instruction-tuned versions to meet the needs of foundational modeling and conversational applications.
• Multi-framework support: Compatible with JAX, PyTorch, TensorFlow, and can also be developed with the Keras toolchain.
• Supports supervised fine-tuning: Suitable for domain adaptation, LoRA fine-tuning, and task customization.
• Multimodal and multilingual expansion: The new generation of Gemma models already emphasizes multimodal understanding and stronger multilingual capabilities.
• Safety and responsible design: Google provides support for data filtering, safety evaluation, and responsible generative AI tools.
• Rich community and ecosystem resources: Models and examples can be obtained through channels such as the official website, Hugging Face, Kaggle, GitHub, and Colab.

Pricing

As an open model family, Gemma itself can be accessed through the official page, Hugging Face, Kaggle, and other channels.
However, its actual usage cost depends on the specific scenario:

• Local use: After downloading the model, you can deploy it yourself, with the main cost coming from local hardware resources.
• Cloud operation: If inference or training is run on Google Cloud or other cloud platforms, compute costs will be incurred.
• Ecosystem tools: Some supporting platforms may provide free quotas or trial resources; refer to the official page for specifics.

Frequently Asked Questions

What is Gemma suitable for?

It is suitable for chat assistants, text generation, code assistance, research experiments, vertical-domain fine-tuning, and AI deployment on local/edge devices.

Is Gemma open source?

Gemma is usually referred to as an "Open Model," providing downloadable weights, but specific use, distribution, and commercial use must still comply with the official license agreement.

Which development frameworks does Gemma support?

Official materials indicate support for JAX, PyTorch, TensorFlow, and provide related implementations and examples.

Where can I get Gemma?

You can access it from the following official entry points:

• Official website: https://ai.google.dev/gemma?hl=zh-cn
• Hugging Face: https://huggingface.co/models?search=google/gemma
• Kaggle: https://www.kaggle.com/models/google/gemma/code
• GitHub (PyTorch implementation): https://github.com/google/gemma_pytorch