Gemma
Google's original open-weight Gemma release, built from the same research as Gemini
Gemma is Google’s first family of open-weight language models, released in February 2024 and built from the same research used inside Gemini. It came out in two sizes, 2B and 7B parameters, each with a base and an instruction-tuned checkpoint, trained on 2 trillion and 6 trillion tokens of web text, math, and code respectively. The context window is a modest 8,192 tokens, in line with other models of that era.
The release mattered because it gave developers a genuinely usable small model they could run on a single consumer GPU or even a laptop, rather than relying on an API. Its license permits commercial use but blocks a list of harmful applications that Google specifies, a middle ground between fully open licenses like Apache 2.0 and the closed terms of most proprietary models. Gemma was later superseded by Gemma 2 and Gemma 3, but it established the naming and licensing pattern Google has followed since.