Human
Mountain View, Calif.: Google has officially released the fourth iteration of its open-weights large language model series, Gemma 4, introducing substantial performance improvements for developers and researchers. Announced today, the new models are built on the same research and technology used to create Google’s flagship Gemini models but are designed to be lightweight and accessible for integration into varied applications.
Gemma 4 is available in two primary model sizes, 9-billion and 27-billion parameters, which Google states are optimized for different hardware configurations, from powerful workstations to resource-constrained environments like mobile devices. The company claims significant advancements in areas such as logical reasoning, coding proficiency, and mathematical problem-solving compared to Gemma 3. Notably, Google asserts that the 27B model outperforms several existing open models in standardized benchmarks.
A key focus of this release is enhanced safety and responsibility, with Google implementing rigorous data filtering and fine-tuning processes to mitigate potential risks associated with generating harmful content. Alongside the model weights, Google is providing an updated Responsible Generative AI Toolkit, offering resources to help developers deploy Gemma 4 ethically. "The Gemma series represents our dedication to open collaboration," a Google spokesperson said. "With Gemma 4, we're empowering innovation across the global AI community."
Unlike proprietary counterparts, the open weights allow users to customize and fine-tune Gemma 4 on their own data, fostering greater flexibility and control over the model's behavior. The models are accessible through platforms such as Kaggle, Hugging Face, and Google’s Vertex AI. Developers can start utilizing Gemma 4 immediately, accelerating development of diverse tools, applications, and research projects.