Google is making further advances in the realm of lightweight AI models — models that can run from your smartphone, a small tablet/PC and similar devices. On that note, Google has now rolled out Gemma 3, the newest addition to its growing AI model family. In addition to this, the company is also rolling out ShieldGemma 2, a 4B-parameter image safety checker designed to detect dangerous content, sexually explicit material, and violence.
“Today, we’re introducing Gemma 3, a collection of lightweight, state-of-the-art open models built from the same research and technology that powers our Gemini 2.0 models. These are our most advanced, portable and responsibly developed open models yet. They are designed to run fast, directly on devices — from phones and laptops to workstations — helping developers create AI applications, wherever people need them. Gemma 3 comes in a range of sizes (1B, 4B, 12B and 27B), allowing you to choose the best model for your specific hardware and performance needs,” Clement Farabet, VP of Research at Google DeepMind, and Tris Warkentin, Director, of Google DeepMind, announced in an official blog post.
This release is part of the company’s ongoing efforts on its AI capabilities, building upon previous models like Gemini and PaliGemma. Gemini, introduced in December 2023, is a multimodal large language model and the successor to Google’s LaMDA and PaLM 2 models, designed to challenge OpenAI’s GPT-4. PaliGemma 2, which was launched a year later, is an upgraded vision-language model capable of more vision-language tasks. Currently, the Gemini family of models than 60,000 Gemma variants in what Google calls the “Gemmaverse.”
Gemma 3 is available in four sizes — 1B, 4B, 12B, and 27B parameters—allowing users to choose a model that fits their hardware and performance needs. According to Google, Gemma 3 outperforms competing models, including Meta’s Llama-405B, DeepSeek-V3, and OpenAI’s o3-mini, on the LMArena leaderboard, particularly in single-accelerator performance.
Interested developers can access the new AI model via AI frameworks like Google AI Studio (which is a free web-based developer tool), Hugging Face, and Kaggle, and the company is also offering $10,000 in AI credits through the Gemma 3 Academic Program (this is for academic researchers only, and applications begin today).
Now, the new model also comes with text and visual reasoning capabilities, allowing it to analyze text, images, and short videos on models with 4B+ parameters (so developers can create apps that can perform the same tasks), which increases its accessibility and usability as well. Additionally, it has an expanded 128k-token context window, significantly improving its ability to process long-form content, such as books, documents, and extended video transcripts. There is also out-of-the-box support for over 35 languages, with pre-trained multilingual capabilities spanning over 140 languages.
Furthermore, developers can leverage function calling and structured output generation (which will enable the creation of AI-driven automation workflows and agent-based experiences) and the 128k-token context window that lets the apps process information to complete tasks. In addition to this, in order to reduce computational requirements while maintaining accuracy, Google is introducing official quantized versions of Gemma 3. These compressed models lower hardware costs and allow for faster performance on resource-limited devices.
