Gemma 4

What it is

Gemma 4 is the fourth generation of Gemma, Google's family of open models — models whose trained weights are published so that anyone can download them, run them on their own hardware and adapt them. It was released on 2 April 2026 and built from the same research and technology as Gemini 3.^[1] The distinction that matters is one of access: a closed model is reached through a provider's interface, while an open-weight model can be inspected, run offline and modified by whoever holds the files.

The four sizes

The generation launched with four variants.^[1] E2B and E4B are small "effective" models of 2 billion and 4 billion parameters, built for places where there is no datacentre at all. The 26B is a mixture-of-experts model (MoE): it holds 26 billion parameters in total but activates about 3.8 billion of them during inference, so the cost of generating each token resembles a much smaller model while the network as a whole is larger. The 31B is a dense model, meaning all of its parameters take part in every token. Between them, the four cover the range from something that fits in a pocket to something that needs a server-class accelerator.^[1]

Context windows depend on size: 128,000 tokens for E2B and E4B, and up to 256,000 tokens for the 26B and 31B models.^[1] A context window is the amount of material a model can hold in view at once — text, and here also image, video and audio tokens. Long windows are what let a model work through a whole document or a long conversation without losing the beginning.

What it runs on

Memory is usually the binding constraint when an open model is deployed. Google's stated figures are that the unquantised 26B and 31B weights fit on a single 80GB NVIDIA H100, while E2B and E4B run offline on phones, a Raspberry Pi and the NVIDIA Jetson Orin Nano.^[1] "Unquantised" means the weights are kept at full precision rather than compressed to a smaller number format, which is the most demanding way to store them; running on a phone or a single-board computer means the small models need far less room than that.

Inputs and languages

All four models process images and video, and E2B and E4B also take native audio input.^[1] Handling several modalities in one model means a single system can be pointed at a photo, a clip or a spoken recording rather than a pipeline of separate specialised components. The models were trained natively on more than 140 languages, a point that matters for communities whose languages are thinly represented in most training data.^[1]

Licence

Gemma 4 was the first Gemma generation released under the Apache 2.0 license; earlier generations used the Gemma Terms of Use.^[1][2] Apache 2.0 is a widely used permissive open-source licence that permits use, modification and redistribution, including commercial use, with few conditions — so the change removes a layer of custom terms that users of earlier Gemma releases had to read and comply with.

Reception and distribution

On the Arena AI text leaderboard for open models, as of 1 April 2026, the 31B model ranked third and the 26B model sixth.^[1] Weights were published on Hugging Face, Kaggle and Ollama, with day-one support in tools such as llama.cpp, vLLM, MLX and LM Studio.^[1] That spread of formats and runtimes is what determines whether an open release is usable in practice: a model that runs in a popular local runtime on the day it appears can be tried without writing new serving code. Google said Gemma models had passed 400 million downloads, with more than 100,000 community variants.^[1]

Where it sits

The family kept growing after launch: a Gemma 4 12B model followed on 3 June 2026, a fifth size sitting between the small effective models and the 26B.^[2] The overall shape of Gemma 4 — a licence change, a spread of sizes from phone to single-GPU, and multimodal input across the range — is what makes it a working option for developers who need to run a model themselves rather than call one over a network.