Google Unveils Gemma 3 AI Model to Compete with Meta's Llama, DeepSeek V3
By
siliconindia | Thursday, 13 March 2025, 12:33 IST

Google has added to its family of open models with the launch of Gemma 3, the newest model in its AI lineup. In a recent blog, the technology giant revealed that Gemma 3 is based on the same technology and research as Gemini 2.0, and comes in 1B, 4B, 12B, and 27B parameter sizes. Google asserts that Gemma 3 is the 'world's best single-accelerator model' that can beat Llama-405B, DeepSeek-V3, and o3-mini models on the LMArena benchmarking platform.
Gemma models are developed for developers and are optimized for speed and can be executed directly on devices such as phones, laptops, and workstations. After the launch of Gemma 1 in February 2024 and Gemma 2 in May, this third version is loaded with improvements that take AI capabilities to the next level.
What's New in Gemma 3?
The new model offers important improvements in reasoning and multimodal processing, processing text, images, and short videos through a 128,000-token context window. It also has in-built support for more than 35 languages and pre-trained support for over 140 languages worldwide.
Important features are:
- AI-driven automation with function calling: Gemma 3 supports function calling and structured output, allowing users to construct smart, responsive applications that automate workflows intelligently.
- Improved performance with optimized models: Official quantized models minimize compute load without loss of accuracy.
- Improved safety features: Google has infused ShieldGemma 2, a 4B image safety scanner that identifies and marks risky content, such as dangerous material, adult content, and violence. The model is compliant with Google's safety guidelines and has been thoroughly tested for safe usage.
- This excited audience who want to experience Gemma 3 can do so by accessing Google AI Studio directly from their browsers, entering a new world of mighty, effective, and safe AI building.