Gemma 3: Google launches new free AI model

Google is making a new version of its open model series Gemma available. Gemma 3 is optimized to run locally on mobile devices, for example. There are four sizes – 1B, 4B, 12B and 27B – these are the billion parameters for each. Google is also releasing ShieldGemma 2, a model that is used for security and can be integrated into applications.

According to Google, Gemma 3 even beats larger models such as Meta's Llama 405B and Deepseek's V3 in some benchmarks. As always, such results are not necessarily conclusive and mean little in terms of actual usefulness. Gemma 3 is pre-trained in 140 different languages and can also use 35 other languages without this training.

The ability to process images, texts and short videos is said to have been improved compared to its predecessor. The context window for these inputs has also been increased to 128 tokens. Gemma 3 allows so-called function calling, which means that the model can use other tools. This allows agentic tasks to be created, i.e., tasks in which the model takes over automatically.

Gemma 3 also comes with quantized versions that reduce the model size and computational requirements while still ensuring a high level of accuracy.

ShieldGemma 2 as a security measure

Google has also released a new version of ShieldGemma. ShieldGemma 2 is a model with four billion parameters that labels images with security concerns. The weights are also open in this model so that users can customize it to their needs. In other words, ShieldGemma is built into an application and warns as soon as images are generated that show sexualized content or violence, for example.

Both models are available via common platforms such as HuggingFace, Google AI Studio and Kaggle. There is a separate program for academics that you can apply for. Those who are accepted get access to Google Cloud Credits.

In the Gemmaverse, as Google calls it, people can share their models and experiences.

(emw)