Google announces Gemma 4 open AI models, switches to Apache 2.0 license

Key Points:

• Google has launched Gemma 4, the latest version of its open-weight AI models, available in four sizes optimized for local use on devices ranging from high-end GPUs to mobile hardware, improving on the older Gemma 3 models.
• The two larger Gemma 4 models (26B Mixture of Experts and 31B Dense) are designed to run efficiently on a single 80GB Nvidia H100 GPU, with reduced latency and higher token processing speeds, while smaller models (E2B and E4B) target mobile devices with low memory and battery usage.
• Gemma 4 supports advanced AI features such as native function calling, structured JSON output, and improved code generation and visual input processing, with context windows up to 256k tokens, enhancing local AI capabilities across more than 140 languages.
• Google has replaced the restrictive custom Gemma license with the permissive Apache 2.0 license to alleviate developer concerns, encouraging broader adoption and innovation within the "Gemmaverse" by granting developers more control over data and deployment.
• The release also confirms an upcoming update to Google's smartphone AI model, Gemini Nano 4, based on Gemma 4 E2B and E4B variants, with developers invited to prototype agentic workflows now for compatibility with this next-gen mobile AI expected to be highlighted at Google I/O.