Google DeepMind · General LLM

Gemma

Google's lightweight open-source model family built from the same research as Gemini, optimized for responsible AI development on accessible hardware.

Overview

Gemma is a family of lightweight open models from Google DeepMind, built on the same technology and research that powers the Gemini models. Available in 2B and 7B sizes (with Gemma 2 offering 9B and 27B variants), Gemma models are designed for responsible AI development with built-in safety features. They achieve impressive performance for their size, particularly in reasoning and coding tasks, and come with comprehensive responsible AI toolkits for developers.

Parameters

2B, 7B (Gemma 1), 9B, 27B (Gemma 2)

Context Window

8K tokens

Architecture

Decoder-only transformer (Gemini technology)

Training Data

6T tokens (primarily English web data)

License

Gemma Terms of Use (permissive, commercial OK)

Capabilities

Efficient text generation for resource-constrained environments

Code generation and technical reasoning

On-device and edge AI deployment

Responsible AI with built-in safety features

Use Cases

Deploying AI models on mobile and edge devices

Building lightweight AI applications with safety requirements

Fine-tuning for specific domains on consumer hardware

Rapid prototyping and experimentation with small models

Pros

  • +Built on Gemini research delivering strong per-parameter performance
  • +Free to use with commercial-friendly license
  • +Small enough for laptop and mobile deployment
  • +Comprehensive responsible AI toolkit included

Cons

  • -Limited context window compared to larger models
  • -Smaller knowledge base than frontier-scale models
  • -Custom license is not standard Apache 2.0 or MIT
  • -Less capable than Gemini for complex multi-step tasks

Pricing

Free and open-source. Runs on consumer GPUs and even CPUs for smaller variants. Available on Google AI Studio for free inference.

Related Models