Google releases Gemma 4: four models with 140 languages and an Apache 2.0 license

Google has released Gemma 4 — a new family of open-source AI models that covers both mobile devices and more powerful server scenarios. The lineup includes four versions right away, support for over 140 languages, and an Apache 2.0 license that removes a significant portion of restrictions for developers and companies.

What Google Showed

Gemma 4 is built on Gemini 3 technologies and comes in four configurations designed for different hardware levels. Google has divided the lineup not only by the number of parameters, but also by architecture type. The junior versions are oriented toward devices with limited resources, including smartphones, while the senior versions are designed for more serious local and server workloads. This is an important move: the company is trying to cover several scenarios with one family, from on-device AI to full-fledged working models for developers.

According to Google, the senior Gemma 4 models demonstrate very high efficiency per parameter. The company cites Arena AI results for text output: the 31-billion parameter model ranked third, and the 26-billion parameter model ranked sixth. At the same time, Google specifically emphasizes that these versions outperformed models that are roughly 20 times larger. Such comparisons should always be read carefully, but the signal is clear: Google is betting not only on scale but also on quality density.

Lineup by Size

Within Gemma 4 there are four different options, each with its own role. For low-power devices, Google offers 2- and 4-billion parameter Effective models. For more serious tasks — a 26-billion parameter Mixture of Experts version and a 31-billion parameter Dense model. The division itself is also indicative: MoE architecture typically activates only part of the network for a specific request, while a dense model uses all parameters. This is not a promise of speed specifically for Gemma 4, but a good hint about how the company differentiates use cases within one lineup.

• 2B Effective — a compact version for devices with strict resource constraints
• 4B Effective — a more powerful option for the same class of local scenarios
• 26B Mixture of Experts — the senior configuration for high-performance systems
• 31B Dense — the largest model in the family
• Support for video, images, and over 140 languages across the entire lineup

Google also added several practical features. All models in the family can work with video and images. The two junior versions also process audio data and understand speech, which makes them particularly interesting for mobile and embedded scenarios. Another important detail is the ability to generate code in offline mode. In practice, this means that Gemma 4 can be used for local development and experimentation with vibe-coding without constant dependence on the cloud and internet connectivity.

Open Apache License

The most notable change for the market is the transition to Apache 2.0. Previous versions of Gemma were distributed under Google's own license, but now the company has chosen a much more familiar and flexible format for the industry.

This simplifies integration into products, provides greater clarity for legal teams, and removes some concerns from teams that want to fine-tune, embed, and commercialize models without extra caveats. Google directly calls such a scheme the basis for "full developer flexibility and digital sovereignty." Behind this formulation stands a quite practical thing: business gains greater control over data, infrastructure, and deployment methods.

Gemma 4 can be tested through Hugging Face, Kaggle, and Ollama — so the entry threshold for experiments and pilots remains low for both enthusiasts and product teams.

What It Means

Gemma 4 looks like Google's attempt to take a stronger position in the open AI segment not through one big model, but through a complete lineup for different devices and budgets. If the stated efficiency is confirmed in practice, and Apache 2.0 truly removes legal barriers, developers will get one of the most convenient sets of open models for local multimodal and code tasks.