Google has launched Gemma 3, an upgraded version of its “open” AI model family. The new models equip developers with powerful AI tools. They run on various devices, from mobile phones to high-performance workstations. Gemma 3 also supports over 35 languages and analyzes text, images, and short videos, making it more versatile.

Enhanced Performance and Hardware Optimization

Google calls Gemma 3 the “world’s best single-accelerator model.” It outperforms Facebook’s Llama, DeepSeek, and OpenAI on a single GPU. The model runs efficiently on Nvidia GPUs and Google’s dedicated AI hardware, improving developer experience.

The vision encoder in Gemma 3 now supports high-resolution and non-square images. Google also introduced ShieldGemma 2, a safety classifier that filters explicit, dangerous, or violent images in both input and output.

Stronger Security and Ethical Standards

Google evaluated Gemma 3’s potential for misuse in creating harmful substances. The company reports a low risk in this area. It focused on STEM-related applications that bad actors could exploit.

Licensing Restrictions and Academic Support

Google still enforces strict licensing rules on Gemma 3. These restrictions limit how users apply the technology. Many question whether Google should label it as truly open-source.

To encourage adoption, Google provides Cloud credits for developers using Gemma 3. The Gemma 3 Academic Program also grants researchers $10,000 in credits for AI research.

Rising Demand for Lightweight AI Models

Developers initially questioned Gemma’s appeal. However, the success of models like DeepSeek proved demand for AI with lower hardware requirements.

With better performance and broader applications, Gemma 3 strengthens Google’s AI presence. However, licensing restrictions continue to spark debate.