Gemma
Google DeepMind's open family of Gemini-derived language models
Gemma is a family of open large language models from Google DeepMind, first released in February 2024 as a lightweight counterpart to Gemini. It has grown through four generations to Gemma 4 in April 2026, with multimodal variants and over 150 million downloads.
- First released February 2024 in 2B and 7B parameter sizes
- Gemma 3 supports over 140 languages and 128K context
- Gemma 4 released 2 April 2026 under Apache 2.0
- Over 150 million downloads and 70,000 variants on Hugging Face
- Gemma 4 31B Dense placed third on Arena's text leaderboard
What it is
Gemma is a series of open large language models built by Google DeepMind, drawing on the same technology as its Gemini models. The first release arrived in February 2024 in two sizes of two and seven billion parameters. Later generations grew in size and capability: Gemma 3 offered 1, 4, 12 and 27 billion parameter versions with support for over 140 languages, and Gemma 4 arrived in April 2026. The family is meant for a wide range of uses, and Google also ships specialised versions for fields such as medicine and coding.
How it came about
Google introduced Gemma in February 2024 as a lighter open-weight counterpart to Gemini. Several outlets read the move as a response to rivals such as Meta publishing their models, and as a departure from Google's habit of keeping its AI code closed. Gemma 2 followed on 27 June 2024, and Gemma 3 on 12 March 2025. On 2 April 2026 Google released Gemma 4 under the Apache 2.0 licence. A 12B version of Gemma 4 followed on 3 June 2026.
Inside the model
Gemma 3 uses a decoder-only transformer with grouped-query attention, a way of sharing attention keys and values to save memory, plus a SigLIP vision encoder for images. Context length is 128K tokens for most models and 32K for the 1B version. Quantized builds trained with quantization-aware training cut memory use substantially, at some cost to accuracy. Gemma 4 comes in Effective 2B, Effective 4B, a 26B mixture-of-experts build and a 31B dense build. Its 12B release uses one unified multimodal design that handles images and audio without separate encoders.
What it can do
Gemma 3 handles text and image input across more than 140 languages. Gemma 4 takes images and video in every size, with native audio input on the E2B and E4B models. On Arena's text leaderboard, the 31B Dense variant placed third and the 26B variant sixth. Smaller Gemma 3n models are tuned to run on phones, laptops and tablets. Quantization reduces memory needs but slightly lowers accuracy and precision.
How to use it
Gemma models are downloadable open weights. Gemma 4 is under the Apache 2.0 licence, while Gemma 3 and earlier versions are source-available under the Gemma Terms of Use. Around 70,000 variants are hosted on Hugging Face, and the family has recorded more than 150 million downloads.
Safety and testing
ShieldGemma 2, a 4B model built on Gemma 3, is designed to spot and filter violent, dangerous and sexually explicit images. Google states that MedGemma is not yet clinical grade.
Reception and debate
Coverage of the first release framed it as a notable opening-up by Google after years of keeping its AI code private, and as a reaction to rivals releasing model code. The family has since been widely adopted, with over 150 million downloads and tens of thousands of community variants on Hugging Face.
Why it matters
Gemma has become a base for specialised models. Google's own variants cover code, translation, machine vision, image safety and medicine, and outside developers build on it too: Tap Health in Gurgaon, India, used MedGemma for AI-assisted diabetes management. Research partners including Georgia Tech and the Wild Dolphin Project worked on DolphinGemma to study dolphin communication.
Questions people ask about Gemma
What is Gemma?
Gemma is a family of open large language models from Google DeepMind, built on technology related to Gemini. It launched in February 2024 as a lighter, openly released counterpart, and has since grown through four generations with many specialised variants.
How is Gemma different from Gemini?
Gemma is based on similar technologies to Gemini but is positioned as a lightweight version. Where Google keeps Gemini as its own product, Gemma weights are published openly: Apache 2.0 for Gemma 4, and source-available terms for earlier releases.
What sizes does Gemma come in?
Gemma 3 was offered at 1, 4, 12 and 27 billion parameters, plus smaller on-device Gemma 3n models. Gemma 4 comes as Effective 2B, Effective 4B, a 26B mixture-of-experts model and a 31B dense model, with a 12B version added in June 2026.
What can Gemma models take as input?
Gemma 3 accepts text and images. Gemma 4 handles images and video in all sizes, and adds native audio input on the E2B and E4B models. The Gemma 4 12B release processes images and audio through a unified architecture with no separate encoders.
Are there specialised Gemma models?
Yes. Google offers CodeGemma for coding, MedGemma for medical analysis, PaliGemma for image work, ShieldGemma 2 for filtering unsafe images, TranslateGemma for translation, and DolphinGemma for research into dolphin communication.
Good reading and sources
- Official websiteOfficial site
- Google Delves Deeper Into Open Source with Launch of Gemma AI ModelBloomberg, Article
- Google Is Giving Away Some of the A.I. That Powers Chatbots (Published 2024)The New York Times, Article
- Welcome Gemma 3: Google's all new multimodal, multilingual, long context open LLMHugging Face, Article
- Gemma — Google DeepMindGoogle DeepMind, Article
The facts on this page come from the Wikipedia article "Gemma (language model)", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.