Google Launches Gemma 3: New Open Models Based on Gemini 2.0

On March 12, 2025, Google unveiled Gemma 3, a significant advancement in AI model technology, incorporating a range of open-weight models that cater to various needs. The release includes models with parameter sizes of 1 billion, 4 billion, 12 billion, and 27 billion. The models aim to deliver enhanced performance while being accessible to developers and researchers.

The larger variants, specifically the 4 billion parameters and above, are notable for their multimodal capabilities, which allow them to process different types of data beyond text. These models provide support for over 140 languages and come equipped with an impressive 128k-token context window, enabling them to handle extensive inputs in a coherent manner. This feature is particularly beneficial for applications requiring deep contextual understanding.

Competitive Performance in Evaluations

In the early phase of LMArena human preference evaluations, Gemma 3 has demonstrated remarkable performance metrics. The 27 billion parameter model has surpassed even considerably larger competitors, such as Llama-405B, DeepSeek-V3, and o3-mini. Such outcomes position Gemma 3 as a formidable contender within the rapidly evolving landscape of AI models, indicating its effectiveness despite being runnable on a single GPU, thereby providing accessibility to a broader range of users and applications.

Open-Door Policy for AI Development

The release of Gemma 3 aligns with a growing trend towards open-source AI models, which fosters innovation and collaboration within the field. By making these models available for public use, Google not only encourages external developers to experiment and build upon this technology but also aims to standardize methods across the industry. As noted by Google in their announcement, the accessibility of models such as Gemma 3 is a step toward democratizing AI capabilities across diverse sectors.

For more details, the official announcement can be read on Google (The Keyword), while further insights can be found in reviews on SiliconANGLE and Hugging Face.