Google releases Gemma 4 under Apache 2.0

A new open release

On 2 April 2026, Google DeepMind released Gemma 4, the newest generation of its family of openly downloadable models. The release was announced in a post by Clement Farabet and Olivier Lacombe ^[1]. Gemma sits alongside Google's closed, API-only models: the weights are published so that anyone can download them, run them on their own hardware, and modify them.

Four sizes, two architectures

Gemma 4 came in four sizes: E2B, E4B, 26B and 31B ^[1]. The two E models are the small "edge" versions, built to run on modest local hardware; the 26B is a mixture-of-experts model and the 31B is a dense one ^[1].

The difference matters for how much computing a model needs. In a mixture-of-experts (MoE) model, the network is split into many separate expert blocks, and each token is routed through only a small subset of them, so the model can hold far more parameters than it computes for any given token. In Gemma 4's 26B model, about 3.8 billion parameters are activated during inference ^[1]. A dense model, by contrast, passes every token through all of its parameters; the 31B version is dense ^[1].

The license change

The most consequential detail is licensing. Gemma 4 moved to the commercially permissive Apache 2.0 license ^[1]. Apache 2.0 is a long-established open-source license that permits use, modification and redistribution of the software, including inside commercial products, as long as the license notice is retained. For a company that wants to build a product on top of a model, that turns a legal review into a formality, which is why license terms often decide whether an open model gets adopted at all.

Benchmarks, downloads and hardware

The 31B model ranked third among open models on the Arena AI text leaderboard ^[1], a leaderboard built from users comparing the outputs of anonymous models side by side. Google also reported more than 400 million Gemma downloads to date ^[1], a measure of how widely the earlier generations were picked up.

Hardware requirements were part of the pitch. The unquantized bfloat16 weights of the 31B model fit on a single 80GB NVIDIA H100 ^[1]. Two terms explain why that sentence is worth stating. Bfloat16 is a 16-bit number format commonly used for model weights; "unquantized" means the weights are stored at that original precision, with no compression applied to shrink them. Because a single H100 is a common unit of GPU capacity, a model whose full-precision weights fit on one such card can be served without splitting it across several machines.

Inputs, context and languages

The E2B and E4B models support native audio input, and all Gemma 4 models accept images and video ^[1], so multimodality is no longer limited to the large versions. The edge models have a 128K-token context window, while the larger models reach up to 256K ^[1]. The context window is the amount of material a model can take into account at once, and with image, video and audio input it covers more than text. Google said the models were natively trained on more than 140 languages ^[1].

Why it matters

Three things define this release. First, the license: Apache 2.0 is a standard, widely understood text, so teams that would previously have negotiated terms can simply ship. Second, the range: the lineup now runs from small edge models with audio input to a dense 31B model that fits on a single high-end GPU, covering phones and laptops as well as servers. Third, the multimodal inputs, which turned up across the entire family rather than only at the top. Against a backdrop of more than 400 million earlier downloads ^[1], Gemma 4 is less a single model than a set of options for people who want to run a capable model themselves, under terms they do not have to ask permission for.