Google Unveils Gemma 4 AI Models: Record Efficiency and Multilingual Mastery
Google has introduced a new generation of open artificial intelligence models, Gemma 4. The lineup is based on Gemini 3 technologies and is distributed under the Apache 2.0 license.
What is known
The series includes four models. Compact versions with 2 and 4 billion parameters are available for mobile and edge devices. More powerful systems feature versions with 26 billion (Mixture of Experts) and 31 billion parameters (Dense). The more there are, the higher the quality of responses, but also greater the computational resource requirements.
Google emphasizes that Gemma 4 demonstrates record efficiency — the ratio of "intelligence to number of parameters." According to the company, the older models have already ranked third and sixth in the Arena AI rankings, surpassing systems that are dozens of times larger in size.
All models support image and video processing and can perform tasks such as optical character recognition. The junior versions additionally work with audio and recognize speech. Another feature is the ability to generate code offline without an internet connection. Overall, the models are trained in more than 140 languages.
Gemma 4 is already available for testing through the platforms Hugging Face, Kaggle, and Ollama.
Source: Google
