Google has launched Gemma 4, a new family of open-weight Artificial Intelligence models that it describes as its most intelligent models so far. The release is aimed at advanced reasoning, code generation and complex logic tasks, and adds multimodal features including built-in audio and visual processing. Google said the family is built using the same research and technology as Gemini 3 and represents a significant step forward for the Gemma series.
The April 2 release builds on growing adoption of the Gemma series, which has seen more than 400 million downloads and 100,000 community-built variants since its debut in February 2024. Gemma 4 is available in four sizes to meet different environmental criteria. The smaller 2-billion- and 4-billion-parameter “Effective” models are intended for edge devices such as smartphones, while the 26-billion-parameter mixture-of-experts and 31-billion-parameter dense models can be deployed in more compute-intensive workloads.
Google positioned the new models as a more efficient option for developers seeking strong performance without the hardware demands of much larger systems. Parameters are the settings a large language model can use to generate an output. While more parameters yield better results, they also require more compute to run. With Gemma 4, Google said it hopes to offer a more effective alternative to previous models, achieving “an unprecedented level of intelligence-per-parameter.” The company also said the models have been trained on more than 140 languages and can run offline.
Google further claimed the larger Gemma 4 models deliver near-frontier performance despite their size, with the 26B variant outperforming models up to 20 times its size. Accessibility is a major part of the release strategy, with Gemma 4 distributed under an Apache 2.0 license, replacing the previous Gemma license. That change gives developers broader freedom to modify and deploy the models commercially across on premises and cloud environments, while maintaining control over data, infrastructure and model deployment.
