Google has announced Gemma 3, an updated version of its “open” AI model family, designed to interpret text, images, and short videos across over 35 languages. According to Google, these models are built for developers creating AI applications that can run on various hardware configurations, from smartphones to workstations.
The tech giant claims Gemma 3 is the “world’s best single-accelerator model,” outperforming competitors like Facebook’s Llama, DeepSeek, and OpenAI when running on a single GPU. The model is optimized for Nvidia GPUs and dedicated AI hardware.
Technical improvements include an upgraded vision encoder supporting high-resolution and non-square images, plus the new ShieldGemma 2 image safety classifier to filter both input and output content for potentially problematic material.
A 26-page technical report provides more details on Google’s performance claims. The release comes as models requiring lower hardware requirements, like DeepSeek, have gained popularity in the AI development community. Despite touting advanced capabilities, Google maintains that Gemma 3 poses a “low risk level” for potential misuse in creating harmful substances.
The definition of what makes an AI model truly “open” or “open source” remains contentious, with Google’s licensing restrictions on usage continuing with this new release. To promote adoption, Google is offering Cloud credits and has announced the Gemma 3 Academic program, allowing researchers to apply for $10,000 in credits to support their work.
Stay tuned to DC Brief for further updates on this story and other technology developments.
