Google Launches Gemma 4 Open AI Models Now

In a milestone for the artificial intelligence ecosystem of artificial intelligence, Google officially unveiled the latest generation of open AI models. This highly anticipated new release from the tech giant is a game-changer for developers, researchers, and companies in general.

Gemma 4: Nature Open AI Models from Google Now not only does it bring native multimodal architectures and performance refreshingly achievable within the commercially permissible Apache 2.0 licence, but more importantly, each model boasts an unprecedented level of intelligence per parameter.

Born of the very same state-of-the-art research and foundational technology that support Google’s in-house Gemini 3, Gemma 4 presents state-of-the-art, frontier-AI capabilities at a fraction of the price. Be it for driving deep reasoning on massive enterprise data centres or natively running offline mode on an Android smartphone, Gemma 4’s arrival signals that open-source AI is synchronised with its closed counterparts at some levels of competition and even slightly ahead in others.

So, what is Google Gemma 4? Four Powerful Configurations

At Google, we understand that 4 different burdens can make equally good rock types. That is why Gemma 4 comes in 4 sizes: Each model is the champion of its respective class:

• Efficient 2B (E2B): Designed strictly for mobile phones and IoT devices by nature, while remaining natively multimodal with strong capabilities. We have engineered it to keep battery consumption below a certain level and RAM usage within acceptable limits.
• The Ultimate Edge Devices: E4B: This model is our greatest portable hitting power analogue. By optimising for edge and mobile use, E4B can boot on such hardware as Raspberry Pi and Nvidia Jetson Orin Nano hardware phones with very little latency indeed.
• Mixture of Experts 26B (MoE): A midrange powerhouse built for low-latency reasoning and compute-heavy tasks suited to workstations or personal systems in an agile enterprise setting.
• Dense 31B: The Gemma family’s flagship. Built for nothing but raw performance, this variant is a leading top performer across the board in every single task we identified on our widely known Arena AI benchmarking. In the open-source model category, Dense 31B currently stands as 3rd worldwide.

By providing this range of models, Google ensures that anyone from a hobbyist coder to a Fortune 500 company can benefit from Gemma 4 without being forced into costly cloud-only infrastructure. nan

Advanced Capabilities: More Than A Simple Chat

While previous large language models (LLMs) were much focused on creating chatterbots, the design philosophy behind Gemma 4 is right for the computer’s next turn: self-consuming workflows and strong reasoning. Here are the key abilities that define this new array of Gemma 4 models:

Strong Reasoning & Multi-Step Planning

Deep logic, mathematics, and instruction-following benchmarks have all received successful breakthroughs that Gemma 4 is capable of. It represents a complex agent capable not only of creating plans for multiple separate stages but also of solving problems too. Outperforms passive text production

Complete Multimodal Understanding

It’s a test model already gone. All Gemma’s 4 models now come standard with video and image input, as well as flexible resolutions for visual tasks like OCR or understanding complex charts. Also, the Edge models E2B and E4B sport native audio input features—providing offline real-time speech recognition and translation in your palm.

Self-Organisation Processes and Offline Code Production

For programmers, Gemma 4 is a silver bullet. It takes care of function calls, structured data output, and system instructions all effortlessly. The models working offline turn a local computer into an AI coding assistant that is entirely private and secure, producing code for offline usage and without consulting any service outside of its own four walls.

Massive Context Windows & Multilingual Fluency

For this generation, Google has completely opened the given passages. This means the E2B and E4B edge models have an entirely jaw-dropping 128K context window, and versions 26B or 31B offer up to 256K. Now it’s possible for programmers to input whole software repositories, lengthy financial documents or even books with but a single prompt. On top of all that, Gemma 4 is natively literate in more than 140 languages — a truly global open AI model if ever there was one.

The Apache 2.0 License: A New Era of Digital Sovereignty

Perhaps the most universally acclaimed aspect of the Gemma 4 launch is that Google has decided to release the entire model family under the Apache License 2.0. The previous versions of Gemma were “open-weight” but came with terms of use specific to Google that restricted certain redistributions and commercial applications.

Through pivoting to the Open Source Initiative (OSI)-approved Apache License 2.0, Google has given the global developer community full autonomy. From here anyone can download, modify, tweak and use Gemma 4 for whatever they care to—whether personal or commercial—without paying royalties.

This move lays the foundations for true digital sovereignty. Businesses that deal with highly confidential data, such as hospitals and financial institutions, can deploy Gemma 4 entirely on premises and keep their own proprietary data strictly within their own secure boundary.

Seamless Enterprise Deployment with Google Cloud

For enterprises who want to deploy these models at scale, Google Cloud has announced that at the same time it is also committed to robust support for the models in Gemma 4. Using Cloud Run and Google Kubernetes Engine (GKE), organisations can efficiently handle demanding inference workloads.

With GKE, teams can leverage the newly updated GKE Agent Sandbox to safely execute LLM-generated code such as tool calls and calls for new runs within their own highly isolated and secure environment. When combined with Nvidia RTX PRO 6000 (Blackwell) GPUs as well as Google’s own TPU accelerators, the cloud infrastructure can enable these models to easily scale from nothing into massive peak loads, meaning that capacity is both used in the best possible way and is very cost-effective.

Expanding the “Gemmaverse”

Since the launch of the first Gemma model, the developer community has downloaded the models more than 400 million times. It has also fine-tuned more than 100,000 versions—a very lively ecosystem Google calls the “Gemmaverse”.

As Google Launches Gemma 4 OpenAI Models Now, the pace is set to pick up markedly further. A combination of the latest architecture of Gemini 3 and the true freedom of the Apache 2.0 License has succeeded in providing Google with what it has now achieved with Gemini 4: the most capable, versatile and developer-friendly open AI model on the market today. AI.ON.iCity Whether you are building the next generation of self-driving car AI or trying to run offline intelligence on a Raspberry Pi, the future has just arrived for open-source AI.

Ready to Get Started? Access Gemma 4 now via Google AI Studio, Hugging Face, Kaggle, and Ollama.