Google Unveils Gemma 3: A New Benchmark for AI Accessibility

Google launched Gemma 3, a collection of open models the tech giant claims are its most advanced, portable, and “responsibly developed” open model yet.

In a published Google Blog post, the company revealed that the open models are designed to run fast and directly on phones, laptops, and workstations, helping developers create artificial intelligence applications wherever and whenever needed. 

Into the Gemmaverse

2025 is the rise and growth of the Gemmaverse: The AI model is engineered on the same foundation as Gemini 2.0. It offers a portable, lightweight, and adaptable developer solution. 

The release coincides with its first anniversary, marking a year of quick adoption (60,000+ community-built variations and 100M+ downloads).

Features

• Parameter sizes: 1B, 4B, 12B, and 27B - to ensure adaptability for different hardware setups

• High performance: Despite being on single configurations, Gemma 3 delivers.

• Seamless workflow automation with function calling 

• Supports over 140 languages to expand global reach

• Enhanced AI applications with advanced text, video, and image analysis

• Balances efficiency and power for resource-constrained environments and mobile (optimized, quantized models)

• Supports large-scale content processing with an expanded 128k-token context window

• Outperformed leading AI models on the leaderboard of LMArena - beating competitors including o3-mini, Llama-405B, and DeepSeek-V3

Benchmarking Excellence

Gemma 3’s 27B outperforms leading chatbots - with a 1338 ELO score despite using only one NVIDIA H100 GPU. 

What’s even better? It offers seamless hardware support and integration, designed to fit into workflows effortlessly and to be compatible across…. 

• Hardware: AMD (via ROCm), NVIDIA GPUs (Jetson Nano to Blackwell), CPUs (Gemma.cpp support)

• Different AI libraries, such as  Google AI Edge, PyTorch, JAX, and Hugging Face

• Kaggle, Colab, and Google Vertex (deployment platforms)

The tech company prioritizes safety by applying different governance policies and STEM-focused risk mitigation.  

ShieldGemma 2 (a 4B-parameter image safety checker) filters violence, explicit material, and dangerous content. The checker was built on the Gemma 3 foundation and launched alongside it. Meanwhile, developers can customize the filter to fit their users' safety needs. 

More:  Google is also introducing its Gemma 3 Academic Program, which offers $10,000 in Google Cloud credits to researchers working on artificial intelligence projects. 

Gemma 3 is poised to be the go-to AI model for developers globally - for its commitment to responsible AI, seamless integration, and robust capabilities. 

Image credit: Google