Google has announced Gemma 3, the latest update to its AI model lineup, and it's making some bold claims. According to the company, this is the most powerful AI model you can run on a single GPU.
A year ago, Google released the first two Gemma models, designed to be lightweight versions of its Gemini AI. Gemma 3 continues in that direction, supporting over 35 languages with pre-trained support for over 140 and working across different platforms, from mobile devices to high-performance workstations.
Google is calling it the “world’s best single-accelerator model,” claiming it outperforms rivals like Facebook’s Llama, DeepSeek, and OpenAI when running on just one GPU. Performance tests on LMArena’s leaderboard show it surpasses competitors like Llama-405B, DeepSeek-V3, and o3-mini.
It has been optimized for Nvidia’s GPUs and dedicated AI hardware, making it particularly efficient for researchers and developers who want high performance without high infrastructure costs. One highlight is a ranking from Chatbot Arena, showing Gemma 3 achieving a high Elo score while only requiring a single Nvidia H100 GPU.
One of the key upgrades in Gemma 3 is its expanded 128K-token context window, allowing it to process and understand vast amounts of information. Its vision encoder now supports high-resolution images and non-square formats, making it more adaptable to various visual processing tasks. For those wanting to dig deeper into the model’s capabilities, Google has released a technical report.
Google has also introduced ShieldGemma 2, a new image safety classifier designed to filter out explicit, dangerous, or violent content. That’s an important addition, especially as AI-generated images become more common.
To encourage the adoption of Gemma, Google continues to offer Google Cloud credits to developers, while academic researchers can apply for $10,000 in credits through the Gemma 3 Academic program. The application form opens today and will remain open for four weeks.
0 Comments - Add comment