Google Gemma

Gemma is Google's family of lightweight, open-source language models built on the same research and technology that underpins the Gemini model family. The models provide developers and researchers with access to Google's AI research advances in a form that can be run on consumer hardware or deployed on smaller cloud instances.

Gemma models are available in 2B and 7B parameter sizes, providing options from an extremely efficient model that runs on most consumer hardware to a more capable model that still fits on typical developer workstations with GPUs. This range of model sizes makes Gemma accessible for different resource constraints.

The models demonstrate that Google's investment in Gemini training and architecture produces improvements that extend to smaller models — Gemma's benchmark performance relative to its size is notably strong, outperforming earlier models of comparable size on many reasoning and language understanding tasks.

Gemma's open-source release includes model weights in both the base pretrained version and instruction-tuned version. The instruction-tuned version is useful out of the box for following user instructions, while the base model is better suited for fine-tuning on specific tasks or domains.

For researchers and developers interested in running Google's AI technology locally, studying the model's capabilities, or building specialized applications without cloud API dependencies, Gemma provides an accessible entry point to Google-quality language AI on consumer-accessible hardware.