Origin
Google’s open-weight Gemma models come from Google DeepMind. According to Wikipedia, DeepMind was founded in the UK in 2010, acquired by Google in 2014, and merged with Google Brain in April 2023 to form Google DeepMind. Google itself was founded in 1998 by Larry Page and Sergey Brin, is headquartered in Mountain View, California, and has been a subsidiary of Alphabet since a 2015 reorganisation.
Wikipedia says Google DeepMind develops two families of language models: the proprietary Gemini and the open-weight Gemma, a lighter family built with technology similar to Gemini’s.
How it got here
Gemma 1 came out on 21 February 2024 in 2B and 7B sizes. Gemma 2 arrived on 27 June 2024 with 9B and 27B models, plus a 2B model on 31 July 2024. Gemma 3 arrived in 2025 in 1B, 4B, 12B and 27B sizes, according to the release notes, and a 270M size, the Gemma 3n edge models and several variants came later that year.
The release notes list special-purpose releases as well, including CodeGemma, PaliGemma, MedGemma and EmbeddingGemma.
Gemma 4 was released on 31 March 2026. Its 31B model card lists five sizes: E2B, E4B, 12B, 26B A4B (a mixture-of-experts model) and 31B. All of them take text and image input. Only E2B, E4B and 12B take audio. The small models have a 128K context window and the larger ones 256K.
Licensing pattern
Google changed the licence with Gemma 4, so one family name now covers two sets of rights.
Gemma 1, 2 and 3 use the Gemma Terms of Use, a custom licence. You may use, modify and distribute the models if you meet its conditions: pass the usage restrictions on to anyone you distribute the model to, give them a copy of the terms, mark files you modified and include a notice file. Wikipedia calls these models source-available rather than open source. The terms page last changed on 1 April 2026 and points Gemma 4 users to a separate licence.
That licence is Apache 2.0. Google’s Gemma 4 license page is the standard Apache text, and the gemma-4-31B-it model card lists Apache 2.0. Unlike the Gemma terms, Apache 2.0 does not make you pass usage restrictions on to anyone you distribute a model or derivative to. Check the generation before you build on it.