The tech giant claims the model outperforms Meta’s Llama and DeepSeek V3, making it a significant leap in lightweight AI development.
Gemma 3, built on Google’s Gemini 2.0 architecture, can analyze text, images, and short videos, making it versatile for various applications. Notably, it is optimized to run efficiently across devices—from smartphones and laptops to high-powered workstations—without requiring extensive computing resources.
Available in 1B, 4B, 12B, and 27B parameter sizes, Google calls Gemma 3 its “most advanced, portable, and responsibly developed open model yet.” The company also claims it is the “world’s best single-accelerator model”, optimized for Nvidia GPUs and AI hardware, ensuring faster and more efficient deployment.
The model comes with an upgraded vision encoder, allowing it to process high-resolution and non-square images alongside ShieldGemma 2, an advanced safety classifier that filters out explicit, dangerous, or violent content from image inputs and outputs.
Gemma 3 supports 35 languages out of the box and boasts pre-trained capabilities in over 140 languages, making it a truly global AI solution. Its 128k-token context window allows applications to process vast amounts of information, enabling complex analysis and interactive intelligence in text and video.
