Google has updated its family of «open» Gemma AI models and introduced Gemma 3 — the most powerful AI to date, according to the company, that can be run on a single graphics card.
In the press release, Google notes that these models are primarily intended for developers who create AI programs that can run on any device (from a smartphone to a workstation). The models support more than 35 languages (including Ukrainian) and can analyze text, images, and short videos.
Gemma 3, according to Google, outperforms competitors such as Meta’s Llama, DeepSeek, and OpenAI when running on a single GPU, and is optimized to work with Nvidia accelerators and other special AI hardware.
The visualization coder has also been updated, so Gemma 3 now supports high-resolution images and non-square images; there is also a new classifier that filters incoming and outgoing images from sexually explicit, dangerous or violent content.
The new Google model is available in four sizes: from 1 to 27 billion parameters, and with a larger context window of 128 thousand tokens (Gemma 2 had a context window of 80 KB), Gemma 3 can understand more information and solve more complex queries.
The models are already available on Hugging Face and online via AI Studio.
Last year, at the time of the Gemma AI family presentation, it was not clear whether it would impress customers — however, with the debut of DeepSeek and similar models, it turned out that there are interest in AI technology with lower hardware requirements.
Source: The Verge, Venture Beat