Google DeepMind has officially launched Google Gemma 4, the latest generation of its open-weight artificial intelligence models. In a major shift for the technology giant, the entire Gemma 4 family is now available under the permissive, industry-standard Apache 2.0 license. This move removes previous restrictions, granting developers and enterprises unprecedented freedom to modify, reuse, and deploy the models for commercial applications.
With four distinct model sizes ranging from edge-optimized versions to a massive 31-billion parameter powerhouse, Google aims to provide flexible solutions for everything from mobile devices to server-scale infrastructure. The release emphasizes advanced reasoning, multimodal inputs, and agent-based workflows, signaling a significant step forward for the open-source artificial intelligence community.
The Shift to the Apache 2.0 License
A defining feature of the Google Gemma 4 release is its transition to the Open Source Initiative approved Apache 2.0 license. Earlier iterations of the Gemma family operated under a custom license that imposed strict limitations on modification, redistribution, and commercial use. The previous license also enforced specific acceptable-use policies and content restrictions.
By adopting Apache 2.0, Google has effectively removed these hurdles. The new licensing structure means developers no longer face monthly active user limits or require a separate agreement with Google to build commercial AI deployments. This standardized approach provides critical licensing clarity, ensuring that organizations have full freedom to integrate these tools into their sovereign and commercial products.
According to Google, this update makes the models significantly more suitable for enterprise and developer applications. The permissive terms of the Apache 2.0 license are the same as those used by other prominent open models, which simplifies the legal landscape for companies looking to build cutting-edge software without restrictive enterprise contracts.
Four Distinct Model Sizes and Hardware Targets
The Google Gemma 4 lineup includes four distinct configurations tailored for different hardware requirements and operational use cases. At the smaller end, Google introduced the E2B and E4B models. These lightweight versions are specifically optimized for edge devices, targeting mobile hardware and low-power environments. While both serve similar devices, the E4B model offers a higher capacity for more demanding local computing tasks.
For heavy-duty, server-scale inference, Google released two massive models. The first is a 26-billion parameter mixture of experts model. This version is designed for complex server-scale operations and currently holds the sixth position on the Arena AI text leaderboard among open models. The second is a 31-billion parameter dense model, which ranks third on the same leaderboard.
Google reports that the 26-billion and 31-billion parameter models can outperform smaller competing models by up to 20 times on the Arena AI benchmark. These performance claims are based on the company’s own evaluations conducted at the time of the models’ release.
Multimodal Capabilities and Agent Workflows
Beyond raw processing power, all four Google Gemma 4 models are equipped with native multimodal support for images and video at various resolutions. Google highlights optical character recognition and chart understanding as key use cases for these visual capabilities. Additionally, the edge-optimized E2B and E4B models feature native audio input, enabling direct speech recognition directly on mobile devices without needing to send data to a cloud server.
The models also boast impressive context windows, allowing them to process large amounts of information in a single prompt. The E2B and E4B edge models support up to 128,000 tokens, while the larger 26-billion and 31-billion parameter models can handle up to 256,000 tokens.
To help developers build complex, automated applications, the entire family supports function calling, structured JSON output, and native system instructions designed specifically for agent-based workflows. Furthermore, Google confirmed that the models were natively trained on over 140 languages, ensuring broad global utility for international user bases.
Expanding the Global Gemmaverse
Since the initial launch of the Gemma family, the developer community has downloaded the models more than 400 million times. This massive adoption has led to the creation of over 100,000 unique model variants, an ecosystem that Google and its community refer to as the “Gemmaverse.” The introduction of Google Gemma 4 under a truly open license is expected to accelerate this growth even further, encouraging more research breakthroughs and real-world applications.
While all four Google Gemma 4 models are fully available for developers to download and use today, some consumer integration details remain unclear. Google has not yet provided a specific timeline regarding when the E2B and E4B edge models might be officially integrated into Android operating systems or other consumer hardware products. Nonetheless, the widespread availability of these models marks a major milestone for open-source AI development.
