Google launches Gemma 4, its most advanced open source AI model family

Google has launched Gemma 4, the new generation of its open source artificial intelligence model family, developed at Google DeepMind using the same technological foundation as the Gemini 3 models. The company offers four variants tailored to different environments: E2B and E4B, designed for mobile devices with low resource consumption; and the 26B and 31B parameter models, aimed at workstations, servers and research settings.

In independent industry rankings, the 31B model holds third place among all open source models worldwide according to the Arena AI leaderboard, while the 26B model ranks sixth. The latter uses a Mixture of Experts (MoE) architecture that activates only 3.8 billion of its parameters during processing, reducing computational cost.

Key technical capabilities include multi-step reasoning, code generation, native support for images, video and audio, and long-document processing through a context window of up to 256,000 tokens in the large models and 128,000 in the smaller ones. The models have been trained on more than 140 languages.

The E2B and E4B models are designed to run fully offline on mobile phones, Raspberry Pi and other low-power devices, developed in collaboration with chipmakers Qualcomm and MediaTek.

The entire family is released under the Apache 2.0 license, which allows unrestricted commercial use. Model weights are available on platforms such as Hugging Face, Kaggle and Ollama, and can be tested directly in Google AI Studio. Since the launch of the first Gemma generation, the models have accumulated over 400 million downloads, with more than 100,000 community-built variants.

Documented applications include the creation of the Bulgarian language model BgGPT by the INSAIT institute, and a collaboration with Yale University to identify new pathways in cancer research.

Key points

• Google launches Gemma 4, its new family of open source AI models.
• Available in four sizes: E2B, E4B, 26B and 31B parameters.
• The 31B model ranks third among open models worldwide according to Arena AI.
• Small models (E2B and E4B) run offline on mobile and low-power devices.
• Context window of up to 256,000 tokens for large models and 128,000 for small ones.
• Trained on more than 140 languages.
• Released under the Apache 2.0 license, allowing free commercial use.
• Available on Hugging Face, Kaggle, Ollama and Google AI Studio.
• The Gemma family has accumulated over 400 million downloads since its first version.