What is Gemma? Gemma is Google’s family of open AI models that you can run locally on your own computer. These lightweight models are built with the same technology as Gemini, but as open models whose model weights you can download free of charge. For professionals who care about privacy or want to work independently of cloud services, Gemma offers some interesting options.
What is Gemma and what sets it apart?
Gemma is a family of lightweight open AI models built with the same research techniques and technology as the Gemini models. It was developed by Google DeepMind and takes its name from the Latin word for gemstone.
The big difference from Gemini is that Gemma is an open model (see open source vs closed AI models). You can download the model weights and use them free of charge, for both personal and commercial purposes. That means no monthly fees and full control over how you deploy the model.
Google has several families of AI models, and Gemma stands out because it is designed specifically to run locally on laptops, workstations and even smartphones.
Which Gemma versions are there?
The first version was released in February 2024, followed by Gemma 2 in June 2024 and Gemma 3 in March 2025. The latest generation is Gemma 4, which comes in five sizes (E2B, E4B, 12B, 31B and the mixture-of-experts model 26B A4B). It accepts text and image input (E2B, E4B and 12B also take video and audio) and has a context window of up to 256,000 tokens. Google now lists Gemma 3 among its previous Gemma models.
Gemma 3 models are available in five parameter sizes: 270M, 1B, 4B, 12B and 27B. The smaller models suit simple tasks such as text classification, while the larger models can handle complex reasoning tasks.
Gemma 3 can handle a wide range of tasks with text and image input, supports more than 140 languages and has a context window of 128,000 tokens. That makes it a good fit for applications that serve a global audience.
What are the benefits of an open model like Gemma?
Gemma is often described as the ‘lightweight’ version of Gemini, but one that is fully accessible to everyone. Its open nature brings some practical benefits.
You can adapt the model fully to your needs through fine-tuning. Google allows commercial use of Gemma without licence fees. Gemma 4 is released under the Apache 2.0 licence; Gemma 3 and earlier versions are covered by the Gemma Terms of Use and Google’s Prohibited Use Policy.
According to Google, Gemma models had passed one billion downloads by 20 August 2026, and more than a hundred thousand Gemma variants had been published in two years. These figures show how popular the models are in the developer community.
An open model offers transparency. Because the model weights are public, you can test the model yourself and assess any safety risks.
What can you use Gemma for?
Developers use Gemma to build AI applications such as chatbots, text summarisation tools and retrieval-augmented generation systems. The range of uses is broad and practical.
Marketing teams use it to generate content, HR departments to screen CVs and consultants to analyse client data. Thanks to its large vocabulary of 256,000 tokens, the model can handle specific and rare terms, which makes it a good basis for further fine-tuning in specialist sectors.
Privacy is a major advantage. Because the model can run locally on a device, sensitive data does not have to be sent to the cloud. That makes it a good option for companies that take privacy seriously.
How well does Gemma perform?
On the LMArena leaderboard, Gemma 3 27B achieved an Elo score of 1338, ahead of DeepSeek-V3 at 1318 and o3-mini at 1304. The difference in model size is large: Gemma 3 reaches this score with just 27B parameters, compared with 671B for DeepSeek-V3.
Tests on a Pixel 9 Pro showed that the optimised model used only 0.75% of the battery for 25 conversations, making it the most energy-efficient Gemma model. That makes it suitable for mobile applications.
The smaller models are fast: the 1B variant processes 2,585 tokens per second during prefill. When you choose an AI model, real-world performance for your own use tells you more than its size.
How do you download and use Gemma?
The latest generation is Gemma 4, with spin-offs such as DiffusionGemma (June 2026) and EmbeddingGemma 2 (October 2026). You can download Gemma 3 models from Kaggle and Hugging Face, and they are available through Ollama, LM Studio and Docker, much like Meta’s Llama. For beginners, Google offers ready-made notebooks in Google Colab.
Deployment is simpler than many professionals expect. Experienced developers can integrate the model straight into their existing workflows. You can run larger models such as Gemma 3 27B locally on consumer GPUs like the NVIDIA RTX 3090.
This makes advanced AI accessible without expensive cloud infrastructure. You decide how and where you use the model, without depending on external APIs or cloud services. If you want to use a local model like Gemma within a team, you need to set down which tasks the model gets and which checks its output has to pass; the vendor-neutral AI Literacy Course covers how to do that.
Frequently asked questions about Gemma
Once you start using Gemma yourself, questions soon come up about the memory you need and the names of the models.
How much memory do you need to run Gemma 4 locally?
According to Google, running Gemma 4 locally takes between 2.9 GB and 69.9 GB of GPU memory, depending on the model size and precision. The smallest model, E2B, fits in 2.9 GB at 4-bit, while the 31B model needs 17.5 GB at 4-bit and 69.9 GB at 16-bit. These figures cover only the model weights, so a long conversation or document takes extra memory for the context window.
Can you use Gemma without installing it yourself?
Yes, you can also use Gemma through the Gemini API without installing anything; according to Google, the API offers the Gemma 4 31B and Gemma 4 26B A4B models. You need an API key for this, which you create in Google AI Studio. The model then runs on Google’s servers instead of your own computer, so you lose the privacy benefit of running it locally.
What does the E in Gemma E2B and E4B stand for?
The E in Gemma E2B and E4B stands for effective, because for these models Google states the effective number of parameters, roughly 2 and 4 billion. Google built these two smallest Gemma 4 models for smartphones, edge devices and the browser. Memory use is higher than the name suggests: according to Google, E2B needs 11.4 GB at 16-bit.
Which specialised Gemma models are there?
Alongside the standard Gemma 4 models, Google offers specialised Gemma models for well-defined tasks. ShieldGemma 2 checks the input and output of generative AI models against predefined safety policies. EmbeddingGemma 2 turns text, images, audio and video into numbers, which you can use for semantic search and retrieval-augmented generation.
