Presenting Google Gemma 3, the most powerful model that can be used with just one GPU or TPU.
It dedication to enabling the application of practical AI technology is based on the Gemma family of open models. As it commemorated Gemma’s first birthday last month, they saw its amazing adoption more than 100 million downloads and its thriving community, which has produced more than 60,000 Gemma variations. This Gemmaverse never stops motivating us.
Google Gemma 3, a suite of cutting-edge, lightweight open models developed using the same technology and research it as Gemini 2.0 models. These are the most cutting-edge, transportable, and ethically created open models to date. No matter where people need them, they are made to run quickly and immediately on gadgets like phones, laptops, and workstations, assisting developers in creating AI applications. Because Gemma 3 is available in a variety of sizes (1B, 4B, 12B, and 27B), you can select the model that best suits your hardware and performance requirements.
Discuss the features of Google Gemma 3, present ShieldGemma 2, and explain how to become a part of the growing Gemmaverse in this post.
Gemma 3 offers more capabilities
- Build using the top single-accelerator model available: In early human preference tests on LMArena’s scoreboard, Gemma 3 outperformed Llama-405B, DeepSeek-V3, and o3-mini, demonstrating state-of-the-art performance for its size. This enables you to design captivating user interfaces that are compatible with a single GPU or TPU host.
- Expand internationally in 140 languages by creating apps that can communicate with your clients in their native tongue. Google Gemma 3 has pretrained support for more than 140 languages and out-of-the-box support for more than 35 languages.
- Develop AI with sophisticated visual and textual reasoning skills: apps that analyse text, photos, and brief videos can be easily created, creating new opportunities for intelligent and interactive apps.
- Use an extended context window to handle difficult tasks: Gemma 3 provides a context window with 128k tokens to enable your apps to process and comprehend large volumes of data.
- Create workflows powered by AI by calling functions: Function calling and structured output are features that Gemma 3 offers to help you automate processes and create agentic experiences.
- Faster delivery of excellent performance with quantised models: Gemma 3 presents official quantised versions, which preserve high accuracy while lowering model size and processing demands.
Strict safety procedures to construct Gemma 3 in an ethical manner
The method strikes a balance between safety and creativity by adjusting testing intensity to model capabilities since it think open models necessitate rigorous risk assessment. The creation of Google Gemma 3 involved thorough data governance, conformity with the safety policies through fine-tuning, and thorough benchmark assessments. Although extensive testing of more competent models frequently guides are evaluation of less capable ones, Gemma 3’s improved STEM performance prompted particular assessments that concentrated on its potential for abuse in the production of hazardous chemicals; the findings show a low risk level.
The development of risk-proportionate safety strategies will be crucial as industry creates increasingly potent models. Over time, it will keep learning and improving the open model safety procedures.
ShieldGemma 2’s integrated security for picture applications
In addition to Google Gemma 3, Google is introducing ShieldGemma 2, a potent 4B image safety checker that is based on Gemma 3. By producing safety labels for three safety categories violence, sexually explicit content, and dangerous content ShieldGemma 2 offers a ready-made solution for picture safety. ShieldGemma can be further tailored by developers to meet their users’ and safety requirements. In order to encourage responsible AI development, ShieldGemma 2 is open and designed to provide flexibility and control by utilising the effectiveness and performance of the Gemma 3 architecture.
All set to connect with the tools you currently utilise
ShieldGemma 2 and Google Gemma 3 easily fit into your current workflows:
- Use your preferred tools to develop: You can select the finest tools for your project with support for Hugging Face Transformers, Ollama, JAX, Keras, PyTorch, Google AI Edge, UnSloth, vLLM, and Gemma.cpp.
- In only a few seconds, begin experimenting: Gain immediate access to Gemma 3 and start developing immediately. Use Google AI Studio to fully explore its possibilities, or download the models from Hugging Face or Kaggle.
- Tailor Gemma 3 to your unique requirements: A redesigned codebase with recipes for effective inference and fine-tuning is included with Gemma 3. Use your favourite platform to train and modify the model, such as Google Colab, Vertex AI, or even your gaming GPU.
- Choose the deployment method that best suits your application and infrastructure with Gemma 3’s many possibilities, which include Vertex AI, Cloud Run, the Google GenAI API, Iocal environments, and other platforms.
- Experience optimal performance on NVIDIA GPUs: From the newest Blackwell chips to the Jetson Nano, NVIDIA has directly optimised Gemma 3 models to guarantee optimal performance on GPUs of any size. Now available in the NVIDIA API Catalogue, Gemma 3 allows for quick prototyping with only an API call.
- Develop AI more quickly on a variety of hardware platforms: Additionally, Gemma 3 interfaces with AMD GPUs through the open-source ROCm stack and is optimised for Google Cloud TPUs. Gemma.cpp provides a straightforward solution for CPU execution.
A “Gemmaverse” of illustrations and instruments
A massive ecosystem of community-made Gemma models and tools, the Gemmaverse is ready to fuel and stimulate your creativity. For instance, AI Singapore’s SEA-LION v3 facilitates communication throughout Southeast Asia by removing language barriers; INSAIT’s BgGPT is a groundbreaking large language model that is the first of its kind in Bulgaria and shows how Gemma can support a variety of languages; and Nexa AI’s OmniAudio exemplifies the potential of on-device AI by introducing sophisticated audio processing capabilities to commonplace devices.
They are establishing the Google Gemma 3 Academic Program to further encourage academic research achievements. To expedite their Gemma 3-based research, academic researchers can request for Google Cloud credits, which are valued at $10,000 each grant. The application will be available for four weeks starting today. Apply online.
Start using Gemma 3
Google Gemma 3 is the next stage in the continuous effort to democratise access to high-quality AI. Are you prepared to discover Gemma 3? This is where to begin:
Exploration in real time:
- Try Gemma 3 in its entirety right in your browser with Google AI Studio; no setup is required.
- Use Gemma 3 with the Google GenAI SDK and obtain an API key straight from Google AI Studio.
Personalise and construct:
- Get Gemma 3 models on Kaggle, Ollama, or Hugging Face.
- Using Hugging Face’s Transformers library or your chosen development environment, you may quickly adjust and modify the model to meet your specific needs.
Scale and deploy:
- Utilise Vertex AI to commercialise your unique Gemma 3 inventions at scale.
- Use Ollama to run inference on Cloud Run.
- Use the NVIDIA API Catalogue to begin using NVIDIA NIMs.
