Google Unveils Gemma 4 AI Models with Apache 2.0 License

Google has launched Gemma 4, the latest iteration of its open-weight AI models, offering developers more flexibility and power. Unlike its predecessors, Gemma 4 is available under the Apache 2.0 license, addressing previous licensing frustrations. The models come in four sizes, optimized for local use, with two large variants, 26B Mixture of Experts and 31B Dense, designed to run on high-end GPUs like the Nvidia H100. These models can also be quantized for consumer GPUs, balancing performance and accessibility.

The 26B Mixture of Experts model is engineered to activate only a fraction of its parameters during inference, enhancing speed without compromising on performance. Meanwhile, the 31B Dense model prioritizes quality, allowing developers to fine-tune it for specific applications. The smaller models, Effective 2B and Effective 4B, are tailored for mobile devices, boasting low memory usage and near-zero latency, thanks to collaborations with Qualcomm and MediaTek.

Google asserts that Gemma 4 models surpass their predecessors in capability, positioning the 31B variant as a top contender in the open AI model arena. Despite their smaller size compared to leading models like GLM-5 and Kimi 2.5, Gemma 4 models are more cost-effective to run, offering a powerful yet accessible option for developers.