What it is
Gemma is a family of small AI models made by Google DeepMind that you can download and run yourself. It is the downloadable cousin of Gemini. Gemini lives on Google's computers and you reach it through Google's apps or its API, a way for software to talk to other software. Gemma models come as files you can put on your own laptop, phone, or server, where they work without sending anything to Google.
Like other language models, Gemma is trained on huge amounts of text so it can answer questions, write, summarise, translate, and help with code. Newer versions can also look at pictures and listen to audio. Google describes Gemma as built from the same research and technology as Gemini, but the models are much smaller. That is the whole point. A smaller model needs less memory and less computing power, so it fits on ordinary devices. The trade-off is that it knows less and makes more mistakes than the giant models behind the big chatbots.
What you download are the model's weights, the billions of learned numbers that make it work. Parameters explains what those numbers are and how size decides which computers can run a model. Gemma sizes are written in billions of parameters as "B", or millions as "M".
The generations at a glance
The dates below come from Google's own announcements and its Gemma release notes.
- Gemma (21 February 2024). Two sizes, 2B and 7B. Google said they could run on a laptop or desktop.
- Gemma 2 (27 June 2024). 9B and 27B models, followed on 31 July 2024 by a 2B model small enough for many phones and laptops.
- Gemma 3 (12 March 2025). Four sizes, 1B, 4B, 12B, and 27B. All but the smallest could read images as well as text, and Google said the family understood over 140 languages. Google called it the most capable model you could run on a single graphics processor, the kind of chip that does AI's heavy maths. A tiny 270M version followed on 14 August 2025.
- Gemma 3n (previewed 20 May 2025, fully released 26 June 2025). Built for phones, taking in text, images, audio, and video. Its two sizes, E2B and E4B, are named for their "effective" size: Google said they have about 5B and 8B parameters but run in as little as 2 GB and 3 GB of memory.
- Gemma 4 (2 April 2026). Four sizes: E2B and E4B for phones and small devices, a 26B model built as a mixture of experts, and a 31B model. Google said they were built from the same research as Gemini 3. A 12B model followed on 3 June 2026, which Google said runs on a laptop with 16 GB of memory.
In practice, the smallest models (270M to 4B, plus E2B and E4B) run on many recent phones and ordinary laptops. The 12B models want a well-equipped laptop. The 26B to 31B models need a powerful graphics card or a workstation, though compressed versions shrink that. These are rough guides; check the requirements for the exact model and tool you use.
The specialised siblings
Google has also released many versions of Gemma trained for one job. A few examples, all dated from Google's announcements:
- CodeGemma (9 April 2024) for writing and completing computer code.
- PaliGemma (14 May 2024) for describing and answering questions about images.
- ShieldGemma (31 July 2024) a safety checker that flags harmful text, followed by ShieldGemma 2 for images in March 2025. Tools like this are a kind of guardrail.
- DataGemma (12 September 2024) which checks statistics against Google's Data Commons, a large public collection of figures from sources such as the United Nations, to reduce made-up numbers. It is an early example of grounding a model in trusted data.
- MedGemma (20 May 2025) for health and medical text and images. Google says it is a starting point for developers and that its outputs are not meant to guide diagnosis or treatment without independent checking.
- Gemma Scope (31 July 2024) a set of research tools for looking inside Gemma, used in the tour described in Interpretability.
Later additions include EmbeddingGemma for search, VaultGemma trained with privacy protections, and TranslateGemma for translation.
Open weights, and a licence that changed
Gemma 1, 2, and 3 came with Google's own Gemma Terms of Use, not a standard open-source licence. At the time of writing those terms allow commercial use by organisations of any size, and Google claims no rights in what you generate. But they add conditions that open-source licences don't:
- A prohibited use policy that lists things you may not use Gemma for, and which you must pass on to anyone you share the model with.
- A wide definition of "derivatives". Any model trained on Gemma's output, for example by using Gemma to write synthetic data for training, counts as a Gemma derivative and falls under the same terms.
- A remote-restriction clause. Google "reserves the right to restrict (remotely or otherwise)" use it reasonably believes breaks the terms.
That is why Gemma was usually described as open weights rather than open source. You got the model files, but not the training data, and not the freedoms a standard licence gives.
Gemma 4 changed this. Google released it under Apache 2.0, a widely used open-source licence, and called it the first Gemma release under a licence approved by the Open Source Initiative, the body that maintains the Open Source Definition. Google said the reason was clarity about developers' rights. The older models, and variants built on them, still come under the Gemma Terms of Use, and some specialised siblings have their own terms. Check the licence file for the exact model you use.
Why it matters
It puts a capable AI on your own device. A model that runs on your phone works on a plane, in a basement, or anywhere without signal. Your text, photos, and recordings stay on the device, which matters for privacy. There is no subscription and no monthly limit. For small businesses handling client data, and for anyone who simply prefers not to send everything to a big company, that's a real difference.
It is Google's answer to the open-model race. Meta's Llama made downloadable models popular in 2023. Gemma arrived in 2024 as Google's entry, alongside Mistral, Qwen, DeepSeek, and Microsoft's Phi, another family of small models. For readers, the competition means more free choices, and each company claims to lead on some measure. Treat those claims with care: launch-day scores are chosen by the company releasing the model, as explained in Benchmarks.
It became a base for thousands of other models. Google calls the community around Gemma the "Gemmaverse". Anyone can take Gemma and adjust it with fine-tuning for a language, a profession, or a single task. Google's own figures show how fast this grew: over 160 million downloads by June 2025, over 200 million by August 2025, and over 400 million by April 2026, with more than 100,000 community variants. Those are Google's counts, and one person can download a model many times, but the scale is clear.
It is small on purpose. Not every job needs the biggest model. Sorting emails, pulling names out of a document, or checking the tone of a message can be done well by a small model that is fast and cheap to run, which cuts the cost of inference. Google designed the 270M model specifically to be fine-tuned for narrow tasks like these, and said in August 2025 that 25 conversations with it used under 1% of a Pixel 9 Pro's battery in its own test.
It shows that openness can grow as well as shrink. Some companies have become more guarded about releasing their models. Google went the other way with Gemma 4, moving from its own terms to a standard licence. Either direction is a business decision that can change again.
Real-world examples
The Gemma model pulled from AI Studio (November 2025)
AI Studio is Google's website where developers try out its models. In late 2025 it offered Gemma alongside Gemini. On 30 October 2025, US Senator Marsha Blackburn, a Republican from Tennessee, wrote to Google's chief executive, Sundar Pichai. She said that when Gemma was asked whether she had been accused of rape, it produced a detailed false story about a 1987 state senate campaign, including an alleged accuser and links to news articles. She wrote that none of it was true, not even the year of the campaign, which was 1998, and that the links led to error pages or unrelated stories. Her letter argued this was not a harmless "hallucination" but defamation. She also cited a lawsuit by the conservative activist Robby Starbuck, who said Google's AI models, including Gemma, had made false and damaging claims about him.
On the evening of Friday 31 October 2025, Google said on X that it had seen reports of non-developers using Gemma in AI Studio to ask factual questions, and that Gemma was never meant to be a consumer tool used that way. It removed Gemma from AI Studio while keeping it available to developers through its API. TechCrunch reported the removal on 2 November 2025. When Gemma 4 launched in April 2026, Google again offered its larger models in AI Studio for developers to try.
The episode shows two things. First, a small model asked about a real, named person can invent specifics with complete confidence, including sources that don't exist. That is hallucination, and small models, which store far less knowledge, are especially prone to it. Second, a model built for developers can end up in front of the public, and when it does, the company that made it answers for what it says. If you use a small model, use it for tasks where you supply the facts, and never ask it about real people and trust the answer.
A cancer hypothesis from a Gemma-based model (October 2025)
On 15 October 2025, Google DeepMind and Yale University released Cell2Sentence-Scale 27B, a model built on Gemma and trained to read data about individual human cells. Google said the model had suggested that a drug called silmitasertib, combined with a low dose of interferon, a natural immune signal, could make certain tumours easier for the immune system to see. According to Google, Yale researchers then tested this in lab-grown human cells not used in training, and the combination raised the relevant immune signal by roughly 50%, while the drug alone had no effect.
Google was careful to call this an early step: a hypothesis confirmed in cells, not a treatment, needing much more testing before it could help any patient. It is a good example of what the Gemmaverse means. Because the weights were open, researchers could take a Gemma model and train it on a completely different kind of data, then publish their version for others.
A Bulgarian-first model built on Gemma 2 (late 2024)
INSAIT, a research institute in Sofia, took Gemma 2 and trained it further on around 100 billion words and word-pieces of text, most of it in Bulgarian, while trying to keep its English skills. The resulting BgGPT models, described in a paper in December 2024, came in 27B, 9B, and 2B sizes and were released under the Gemma Terms of Use. Google features BgGPT among its Gemmaverse examples. For smaller languages, which big AI companies often serve poorly, this kind of local adaptation is one of the most practical uses of an open model.
An everyday case: the offline field notes
Picture Tomasz, a surveyor who spends days on rural sites with no phone signal. He installs a free app that runs a small Gemma model on his phone. At the end of each visit he dictates rough notes, and the phone turns them into a tidy summary and a list of follow-ups, entirely offline. He learns its limits fast: it sometimes mishears technical terms, and when he once asked it about local building rules, it confidently quoted a regulation that doesn't exist. So he uses it only to tidy what he has said, and checks anything factual when he is back online.
Try it yourself
This exercise runs Gemma on your phone, then tests it with the internet switched off. It takes about 20 minutes, mostly downloading. There's a laptop option after the steps.
What you need, at the time of writing: an Android phone running Android 12 or later, or an iPhone running iOS 17 or later, with a few gigabytes of free space. A recent phone with plenty of memory works best. App names, requirements, and model choices change often, so check the store listing first, and only install the app published by Google.
- Install the app. Open the Google Play Store or Apple App Store and search for "Google AI Edge Gallery". Check that the developer is Google. At the time of writing it is free, and since April 2026 it no longer needs a separate Hugging Face account to download models.
- Download a Gemma model. Open the app, choose AI Chat, and pick a Gemma model. At the time of writing the app offers Gemma 4 E2B for mid-range phones and E4B for high-end phones. Start with E2B if you're unsure. Use Wi-Fi, as the download is large.
- Turn on flight mode. Once the model has downloaded, switch your phone to flight mode, and check that Wi-Fi is off too. Everything from here runs on the phone.
- Try a practical task. Paste this: > Rewrite this message so it sounds friendly but firm: "Your invoice is 30 days late. Please pay now."
- Give it your own facts. Small models are best when you supply the information. Paste: > Here are my notes from a meeting: "Launch moved to 14 March. Sam does the cover. Budget 800 dollars. Need proofreader by Friday." Turn these into a short email to my team with a to-do list.
- Test its knowledge, and watch for invention. Ask: > Who won the most recent football World Cup, and what are three news stories from this week? It has no internet, and its knowledge stops at its training date. Notice whether it admits that, or makes something up. Then ask about a local business or a person you know well, and see how confidently it invents details.
- Try a photo, if offered. If the app has an image option, take a photo of a menu or a sign and ask: > Describe what is in this picture and translate any text into English.
- Tidy up. Turn flight mode off. Models take up a lot of space, so delete any you no longer need from the app's model list.
On a laptop instead: install the free app Ollama from ollama.com (Windows, Mac, or Linux), open a terminal, and type ollama run gemma3:4b. At the time of writing that download is about 3.3 GB; on an older computer use ollama run gemma3:1b, about 0.8 GB. Gemma 4 is also listed, but its smallest standard download is larger, about 7 GB. Then turn off Wi-Fi and use the same prompts. Parameters explains how to judge whether a model will fit on your machine.
Common mistakes and misunderstandings
Confusing Gemma with Gemini. Gemini runs on Google's servers and is reached through Google's apps. Gemma is a separate family of smaller models you run yourself. They share research, not a login, and a Gemma model on your phone knows much less than Gemini.
Using a small model as a search engine. A 2B or 4B model can't store most of the world's facts. Asked about people, events, laws, or numbers, it will often invent them fluently, as the November 2025 episode showed. Give it the facts and ask it to rewrite, sort, or summarise; check anything it tells you from its own memory.
Assuming every Gemma has the same licence. Gemma 4 uses Apache 2.0. Gemma 1 to 3, and models built from them, come under the Gemma Terms of Use, with a prohibited use policy and conditions that pass on to anything you build or share. Specialised siblings may have their own terms. Read the licence for the exact model you use.
Thinking "runs offline" means "private everywhere". The model on your device doesn't send your text anywhere, but the app around it might collect usage data, and a cloud service that offers Gemma is not local at all. Check the app's privacy settings and terms.
Trusting any download with "Gemma" in the name. Anyone can publish a Gemma variant, and some are fine-tuned by strangers. Download from Google's official pages or well-known tools, and be careful with variants that promise to have their safety rules removed.
Treating a specialist model as a professional. MedGemma is a building block for developers, not a doctor. Google says its outputs need independent checking and are not meant to guide diagnosis or treatment. The same caution applies to any specialised model.
Expecting laptop speed from a phone. Bigger models on older phones can be slow, drain the battery, and make the phone hot. Start with the smallest model and move up only if it copes.
Related elements
- Gemini: Google's large models on its own servers, built from the same research as Gemma.
- Parameters: what model sizes mean and how to tell whether a model will run on your device.
- Llama: Meta's open-weight family, which made downloadable models popular and has a licence worth comparing.
- Phi: Microsoft's small-model family, another option built for modest hardware.
- Mistral: a French company whose open-weight models are also popular for local use.
- Qwen: Alibaba's open-weight family, another common base for community variants.
- Fine-tuning: how the Gemmaverse's thousands of variants are made.
- Hallucination: why small models invent facts, as in the AI Studio episode.
- Interpretability: a hands-on tour that uses Gemma Scope to look inside a Gemma model.
- Privacy: why keeping a model on your own device can protect sensitive data.
- Inference: what it takes to run a model, and why small models are cheaper to run.