On Wednesday, Google unveiled the Gemma 3 family of artificial intelligence (AI) models. This new series follows the Gemma 2 lineup, launched in August 2024, and features advancements in both text and visual reasoning. The technology company, based in Mountain View, announced that these open-source models support over 35 languages, with the potential for fine-tuning to accommodate 140 additional languages. According to Google, the models are designed to operate efficiently on a single GPU or the company’s Tensor processing unit (TPU).
Google Unveils Gemma 3 Series AI Models
In a recent blog post, Google outlined the features of the new small language models (SLMs), which leverage the same technology as the Gemini 2.0 models. Known for their open-source nature and on-device performance, the Gemma series has achieved over 100 million downloads and facilitated the creation of more than 60,000 model variants.
The tech company claims that the Gemma models outperform Meta’s Llama-405B, DeepSeek-V3, and OpenAI’s o3-mini AI models according to the LMArena’s leaderboard. Additionally, these models are offered in four variations — 1B, 4B, 12B, and 27B parameters, and are compatible with single GPU or TPU configurations.
The Gemma 3 series enhances capabilities in both text and visual reasoning by enabling analysis of images, text, and brief videos. These models come with a context window of 128,000 tokens and support function calling, which allows developers to create agentic functionalities in their applications.
Google emphasized that the AI models’ development involved meticulous risk assessment, underpinning their internal safety protocols through fine-tuning and benchmark testing. Moreover, the models underwent evaluations using more advanced prototypes, with the company asserting a low risk level for these models.
Alongside the Gemma 3 series, Google also introduced ShieldGemma 2, a 4B parameter image safety checker aimed at preventing the generation of harmful, sexually explicit, or violent content. Developers are encouraged to customize ShieldGemma further to enhance its safety parameters. The Gemma 3 series of AI models can be accessed via Google’s Hugging Face listing or on Kaggle.
