Gemma 4 FAQ

Gemma 4 is Google DeepMind's family of open AI models, released on March 31, 2026. It comes in five sizes, from a model that runs on a phone to one for workstations and servers, and you can download and run it yourself. Start on the home page.

Yes. The model weights are free under the Apache 2.0 licence, including for commercial use. You pay only for the hardware or hosting you run them on. See pricing.

Five: E2B and E4B for phones and small devices, 12B for laptops, and 26B A4B and 31B for a strong GPU or server. Details on the models page.

Gemini is Google's larger, closed model family that Google runs for you. Gemma is the open family you download, run privately and can modify. Gemini's top models are more capable. See the model comparison.

Google DeepMind. This website is an independent community reference and is not affiliated with Google.

January 2025. For anything more recent, paste the information into your prompt or connect the model to a search tool.

More than 140, according to Google's model card, built on a vocabulary of about 262,000 entries.

Much better. Against Gemma 3 27B, the 31B rises from 20.8 to 89.2 on AIME 2026 maths and from 29.1 to 80.0 on LiveCodeBench v6 coding. It also adds a thinking mode, audio input and a longer context. See the Gemma 3 page.

Pick the largest one your memory allows: E2B or E4B on a phone, 12B on a 16 GB laptop, 26B A4B for speed on a 24 GB GPU, and 31B for the best quality. The runtime selector can help.

The 31B scores highest on almost every test. The 26B A4B is close behind and faster. See benchmarks.

Sometimes. The gap is two to three points on knowledge, maths and coding, but wider on agent tasks (76.9 vs 68.2) and long-context retrieval at 128K (96.4 vs 89.8).

The E stands for "effective": the edge models run with the footprint of a 2B or 4B model while holding more parameters in total. A4B means about 4 billion parameters are active per token in the 26B Mixture-of-Experts model.

It is a Mixture-of-Experts model. Each token uses about 3.8B of its parameters, so there is far less to compute. The whole model still has to fit in memory.

It is the largest model that accepts audio, it has the full 256K context, and it fits a 16 GB laptop at 4-bit. See the 12B guide.

Not in Gemma 4. If you need something tiny, Gemma 3 has a 270M model.

The -it (instruction-tuned) model. Base models continue text instead of answering questions and exist for people doing their own training.

From Hugging Face, Kaggle, Ollama or LM Studio. The download hub has links for every size, and how to download walks through it.

Ollama: install it, then run ollama run gemma4. If you prefer no terminal, use LM Studio. See the install guide.

For running locally, the official QAT 4-bit build of the largest size your memory allows - in Ollama, the -it-qat tag. For serving or fine-tuning, the plain -it repository. See QAT builds.

Through Ollama's QAT tags, from about 4.3 GB for E2B to about 19 GB for the 31B. Full-precision files are several times larger.

A small draft model that works alongside the main model to speed up generation. It is not a chat model and does nothing useful by itself.

They are separate naming schemes. Ollama uses gemma4:12b; Hugging Face uses google/gemma-4-12B-it, with a capital B.

Yes. Once a model is downloaded, it needs no internet connection.

No. You can try it in Google AI Studio, or see the playground and chat pages.

Yes. Apple silicon Macs run it well through Ollama, LM Studio or MLX. See Mac setup.

At 4-bit, Google lists about 2.9 GB for E2B, 4.5 GB for E4B, 6.7 GB for 12B, 14.4 GB for 26B A4B and 17.5 GB for 31B, for the weights alone. Long conversations need more. See the VRAM guide.

Yes. The 12B at 4-bit needs about 6.7 GB for the weights, which leaves room for everything else.

The two edge models are usable on a CPU. Larger models will run but slowly.

Little with Google's QAT builds, which are trained to tolerate 4-bit. Ordinary 4-bit conversions lose more, and below 4-bit quality falls quickly.

In Google's tests, E2B writes about 52 tokens a second on a Galaxy S26 Ultra, 56 on an iPhone 17 Pro and 160 on a MacBook Pro M4 Max, all on GPU.

Multi-Token Prediction pairs the model with a small drafter that proposes several tokens at once for the main model to check. Google describes it as faster with no quality loss. See capabilities.

Yes. The reasoning is generated text, so it costs time and tokens. Turn it off for quick tasks.

Text and images on every size, video as a series of frames on every size, and audio on E2B, E4B and 12B. Output is always text. See features.

No. It reads images and audio but only writes text.

256K tokens on the 12B, 26B A4B and 31B, and 128K on E2B and E4B. Google's phone builds list a 32K maximum.

A setting where the model writes out its reasoning before the final answer. It improves accuracy on maths, code and planning. Published benchmark scores assume it is on.

Yes, with built-in tokens for declaring and calling tools. The model only writes the request; your program runs the tool and should validate it first.

Yes. For small text or dense pages, raise the image detail setting to 560 or 1,120 tokens per image; the default is 280.

Yes on E2B, E4B and 12B. Google's model card limits clips to 30 seconds, so split longer recordings.

Yes. The weights are open and the licence lets you keep and redistribute what you build.

No. It only sees the current conversation. Any memory between chats comes from the app you use.

Yes. Google serves Gemma 4 through the Gemini API, and other providers host it too. See the API page.

No. On Google's API you use a Gemini API key from Google AI Studio. Other providers issue their own keys, and a local server needs none.

Google's documentation lists the 31B and 26B A4B. For the smaller sizes, use another provider or run them yourself.

Google's pricing page lists no per-token price for Gemma models, and access is rate-limited. Other providers charge per token.

On OpenRouter on October 7, 2026, the 31B was listed at $0.09 per million input tokens and $0.34 per million output tokens. Prices change; check pricing.

Yes with local servers such as Ollama and LM Studio and with most third-party providers, which accept the OpenAI format. Google's own examples use the Google GenAI SDK.

Use an API to start quickly or for light use. Host it yourself for privacy, offline use, the smaller models or steady high volume. See self-hosting.

Yes. E2B and E4B are built for it and work fully offline. See mobile downloads.

The official route is Google AI Edge Gallery, a free app from Google on the App Store and Google Play. Gemma 4 is downloaded inside it.

iOS 17 or later, or Android 12 or later, and storage for the model: 2.58 GB for E2B or 3.66 GB for E4B. See the Android page.

GPU. On Google's test phone it cut the wait for the first word from 1.8 seconds to 0.3.

No. There are no mobile builds beyond E4B. The 12B and larger need a computer.

While it is generating, yes. When idle it uses nothing.

The larger sizes are. The 31B solves 80.0% of LiveCodeBench v6 problems and the 12B 72.0%. The edge models score much lower. See Gemma 4 coding.

The 31B for quality, the 26B A4B for speed, and the 12B on a 16 GB laptop.

Yes. Run it with Ollama or LM Studio and point any extension that accepts an OpenAI-compatible endpoint at the local address.

Yes, through Ollama, as an unofficial setup that Anthropic does not support. See Gemma 4 vs Claude Code.

It is the family's weakest area. On the Tau2 agent test the 31B scores 76.9 but E2B only 24.5. Use a 12B or larger and keep task chains short.

Yes. EmbeddingGemma 2, released October 6, 2026, turns text, images, audio and video into vectors for search. See the EmbeddingGemma 2 guide.

Apache 2.0, a standard permissive open-source licence.

Yes. Apache 2.0 allows commercial use, modification and redistribution, including in closed-source products. Keep the licence and copyright notices when you redistribute.

No. Gemma 3 and earlier remain under the Gemma Terms of Use.

When you run the model on your own device, your prompts are not sent to a model provider. Hosted services do receive them, so read their terms.

Yes. Like every language model it can state wrong or outdated things confidently. Check anything important. See limitations.

There is no server-side filter on a model you run yourself. Safety behaviour is trained in, and adding safeguards to a public product is up to you. See safety.

Only with a qualified person checking the output. Fluent answers can sound authoritative when they are wrong.

Update your app and download the model again. A July 2026 refresh of the weights fixed the chat template and tool calling without changing the model name.

Local apps often start with a context window far smaller than the model supports. Raise the context length in your app's settings.

Check the name. Gemma 4 repositories use a capital B, for example google/gemma-4-12B-it.

Most often the model does not fit in memory, or it is running on the CPU instead of the GPU. Try a smaller size or a QAT build, and switch thinking off for simple tasks.

Your app is not separating the thinking from the reply. Update the app, or turn thinking off.

Published scores use thinking mode and, for vision, the highest image detail. Defaults are lower. Quantised builds can also score slightly lower.

For experimenting, yes. For anything that must stay the same, name the size and build explicitly, because a general tag can be repointed.

Try the prompting guide and technical specs, or use the contact page.