Google released 2B and 7B Gemma weights in pretrained and instruction-tuned variants. Developers could start with a base checkpoint for fine-tuning or use an instruction-tuned model directly.
Gemma
Also known as: Google Gemma · Gemma 4
Plain English
Gemma is Google’s family of downloadable foundation models. It includes small text models for local use and larger releases that can accept images or other media.
Technical Overview
Gemma draws on research behind Gemini and includes pretrained and instruction-tuned checkpoints. Gemma 3 combines local and global attention, limiting the amount of cached state required by most layers. Gemma 4 adds dense and sparse variants and publishes weights under Apache 2.0, while previous releases used Gemma’s own terms.
Timeline (4)
2024
Release- Release
Google released Gemma 2 in 9B and 27B sizes. The 27B model was designed to run at full precision on a single supported accelerator, making it possible to serve a larger checkpoint without sharding.
2025
ReleaseGemma 3 brought image input and a 128K context window to its 4B, 12B, and 27B models. Google also made quantized versions available to lower the memory required for local inference.
2026
ReleaseGoogle released Gemma 4 under Apache 2.0. The lineup included the sparse 26B-A4B model, which activates about 4B parameters per token, alongside dense models for different memory budgets. Google reported initial Arena rankings at launch.
See also (8)
Open full graphGemma