Google releases Gemma 3, its most advanced compact AI model

Google has released Gemma 3, the successor to their range of lightweight open models powered by their Gemini 2.0 models. In a blog post making the announcement, Google said Gemma 3 was available in sizes including 1B, 4B, 12B and 27B. The tech company claimed these models are their “most advanced, portable and responsibly developed open models yet.” 

Just like the earlier two versions, the Gemma 3 models have also been built so they can directly run on phones, laptops and devices, and help developers create AI applications. 

The models have support for over 35 languages, and can analyse text, images, and short videos, but respond just in text. 

The company has claimed that Gemma 3 outperforms Meta’s Llama-4-5B model as well as OpenAI’s o3-mini and DeepSeek-V3 on several benchmarks while running on a single GPU.

Gemma 3 also has a 128k-token context window which means the models can process large amounts of information. 

Furthermore, Google has released the ShieldGemma 2 image safety classifier to filter out images that are sexually explicit, dangerous and violent.

The models can be downloaded via platforms like Kaggle and Hugging Face or even Google Studio.