Google said on the 3rd that it unveiled the next-generation version of its open model lineup "Gemma," called "Gemma 4."
The newly announced Gemma 4 is characterized by intelligence that is further improved compared with the transfer version, delivering higher performance than large-scale models in reasoning, code generation, and complex logical processing.
Gemma 4 was developed based on Google's Gemini research and technology, and includes features that support agent-based workflows beyond simple conversational functions. It also has built-in vision and audio processing, enabling offline operation without latency in mobile and Internet of Things (IoT) environments.
In terms of performance, on Arena, a global AI text evaluation leaderboard, the 31B model ranked third among open models and the 26B model ranked sixth.
Gemma 4 is offered in various sizes, including ▲Effective 2B (E2B) ▲4B (E4B) ▲26B mixture of experts (MoE) ▲31B dense, and is designed to let developers use high-performance artificial intelligence (AI) features while reducing hardware burdens.
It also applies the Apache 2.0 license to allow commercial use, and enables developers and corporations to secure control over data infrastructure and deployment methods.
For security, it applies Google DeepMind's safety protocols and is designed to address key risks such as supply-chain vulnerabilities and data poisoning.