Google has officially released the open-source Gemma4 series of models, marking a major advancement in the field of artificial intelligence. The unique feature of these models is that one of them, with only 380 million parameters, can outperform other large models that are 20 times its size on authoritative rankings. This means that powerful AI capabilities can easily be embedded into smartphones and lightweight laptops, providing users with convenient AI services.
The Gemma4 series includes models of various sizes, specifically gemma-4-E2B with an effective parameter count of 2.3B, gemma-4-E4B with 4.5B parameters, a mixture-of-experts model with 26B parameters, and a dense model with 31B parameters. These models enhance computational efficiency and performance through innovations in their underlying architecture. The newly introduced layer-wise embedding technology allows small models to maintain high-speed computation while possessing knowledge reserves beyond their own size.
Additionally, Gemma4 has made significant improvements in attention mechanisms. The model uses a hybrid attention architecture that combines the advantages of local sliding windows and global attention, ensuring efficient memory usage when processing long texts. The application of mobile devices and Internet of Things terminals will be greatly enhanced, with the E2B and E4B models of Gemma4 specifically optimized for computing and memory efficiency, making AI applications on phones feasible.
In multiple industry standard tests, the Gemma4 series delivered satisfactory results. The 31B parameter dense model ranked third globally among open-source models on the Arena AI text leaderboard, while the 26B parameter mixture-of-experts model also achieved sixth place. These models demonstrate strong capabilities in areas such as text generation, mathematical reasoning, and code writing.
To promote developers' use, Google has placed Gemma4 under the Apache 2.0 license, supporting flexible deployment both locally and in the cloud. Gemma4 has also received support from several mainstream platforms, allowing developers to easily download model weights and quickly build intelligent applications using machine learning tools. Google's release not only showcases the potential of top-tier artificial intelligence but also creates a more open and secure environment for developers.
Official blog: https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/
Key points:
🌟 The Gemma4 series of models is open source, with small parameter counts but powerful performance.
📱 Optimization for mobile devices and Internet of Things terminals enables efficient AI applications.
🚀 Achieved excellent results in multiple industry standard tests, promoting developers in building intelligent applications.
