eWEEK content and product recommendations are editorially independent. We may make money when you click on links to our partners. Learn More.
The development team with Google recently lifted the veil on Gemma 3 – the latest entry in the Gemmaverse and the Gemma family of open-source AI tools. Touted as the “world’s best single-accelerator model” to date, Gemma 3 is based on much of the same tech that is seen in the closed-source Gemini 2.0 model.
According to the Chatbot Arena Elo Score, a popular platform for ranking AI chatbots and large language models (LLMs), Gemma 3 outperforms DeepSeek-V3, o3-mini, Llama3-405B, Mistral Large, and Gemma 2.
Gemma 3’s features and functionality
Gemma 3 underwent extensive pre- and post-training processes before being released to the public, including distillation and model merging. Developers also used reinforcement learning based on human feedback, machine feedback, and execution feedback. As a result, Gemma 3 is significantly better at following user instructions, performing mathematics, and generating computer code.
But that’s just the beginning of the new features and functionality seen in Gemma 3. In addition, Gemma 3:
- Supports function calling and structured output for use with AI agents.
- Features a 128k-token context window.
- Uses different algorithms for both text and image processing.
- Is capable of supporting 140+ languages.
- Is available in four sizes: 1B, 4B, 12B, and 27B.
Gemma 3 supports many of the most popular development tools including Gemma JAX Library, Gemma.cpp, llama.cpp, Ollama, and Hugging Face Transformers. This open-source AI model can be deployed in multiple ways, including via Google GenAI API, Cloud Run, Vertex AI, Cloud GPU, or Cloud TPU.
Integrated safety features with ShieldGemma 2
Known as ShieldGemma 2, the new image safety model was trained on various datasets of both natural and computer-generated imagery. When compared to internal benchmarks that include LlavaGuard 7B, GPT-4o mini, and even Gemma 3, ShieldGemma 2 shows the most accurate image labeling results.
ShieldGemma 2 includes a built-in safety protocol that lets it organize images into one of three categories: dangerous content, sexually explicit, and violence. The development team with Google has gone to great lengths to ensure that Gemma 3 maintains alignment with their established data governance and safety policies.
Making AI more accessible with Gemma
The various types of AI models that make up the Gemmaverse are designed to run efficiently on smartphones, laptops, and desktops. With more than 100 million downloads to date, and with more than 60,000 variants comprising the Gemmaverse, Google is well on its way to making AI more accessible.