Gemma 4 Model Overview
Summary
Documentation for Gemma 4, an AI model family with multimodal capabilities, varying sizes, and open weights for commercial use.
Key quotes
"Gemma models are provided with open weights and permit responsible commercial use, allowing you to tune and deploy them in your own projects."
"Gemma 4 model family spans four distinct architectures tailored for specific hardware requirements: Small Sizes, Dense, Mixture-of-Experts, and Unified."
"All models in the family are designed as highly capable reasoners, with configurable thinking modes."
The page provides a detailed technical overview of the Gemma 4 model family, including its four distinct architectures (2B, 4B, 26B MoE, 31B dense, and 12B unified) optimized for mobile to server-grade deployments. It covers key features such as extended multimodal support (text, image, video, and audio), increased context windows, and enhanced coding capabilities. Memory requirements for inference are tabulated for various parameter sizes and quantization approaches, with notes on efficient deployment strategies like speculative decoding and Quantization-Aware Training (QAT). The documentation also references resources for downloading models and earlier Gemma versions.