Google has unveiled the Gemma 3 series of artificial intelligence (AI) models, marking a significant upgrade from the previous Gemma 2 lineup launched in August 2024. These new open-source models feature enhanced text and visual reasoning capabilities, supporting over 35 languages while offering potential fine-tuning for an additional 140 languages. The Mountain View tech company emphasized that the models are optimized for operation on a single GPU or Google’s Tensor Processing Unit (TPU).
Google Releases Gemma 3 Series AI Models
In a recent blog post, Google explained that the small language models (SLMs) in the Gemma series were built upon the same technology foundation as the flagship Gemini 2.0 models. The Gemma series is recognized for its open-source nature and effective performance on devices. According to the company, the Gemma models have been downloaded more than 100 million times and are credited with generating over 60,000 different variants.
Google asserts that the Gemma models surpass Meta’s Llama-405B, DeepSeek-V3, and OpenAI’s o3-mini on the LMArena leaderboard. The new models come in four sizes—1 billion, 4 billion, 12 billion, and 27 billion parameters—and are designed to operate efficiently on a single GPU or TPU.
The Gemma 3 series is equipped with sophisticated text and visual reasoning abilities, enabling it to analyze images, text, and short video content. It features a context window capable of handling up to 128,000 tokens and includes function calling support, which facilitates the development of more interactive applications and software.
Google has assured that the development of these AI models incorporated meticulous risk evaluations, alongside adherence to internal safety policies through fine-tuning and benchmark testing. The company reported that the models showed a low-risk profile during these evaluations.
Alongside the Gemma 3 series, Google also introduced ShieldGemma 2, a 4 billion parameter image safety checker aimed at preventing the creation of harmful, sexually explicit, or violent content. Developers will have the flexibility to further customize ShieldGemma to enhance its safety protocols. The Gemma 3 AI models can be accessed and downloaded from Google’s Hugging Face listing or on Kaggle.
