Gemma 3 by Google

Gemma 3 is a series of open-weight foundation models developed by Google and available in parameter sizes of 2B, 7B, 12B, and 27B. The models support multimodal inputs including text, images, and short video clips, and feature a 128,000-token context window for handling long-form documents and extended conversations. Gemma 3 is designed to run efficiently on a single GPU or compatible consumer hardware, including Apple Silicon and NVIDIA RTX cards. Model weights are available at no cost through Google AI, Hugging Face, and Kaggle under the Apache 2.0 license, which permits both research and commercial deployment. Gemma 3 is used by software developers building local or on-premise AI applications, researchers adapting models for specialized domains through fine-tuning, and organizations that need to run inference without sending data to external cloud services. Common use cases include building multilingual chatbots, integrating vision-language understanding into mobile applications, processing long documents, and deploying AI on edge hardware where latency and connectivity constraints apply. The models support LoRA fine-tuning, making domain adaptation practical without requiring full retraining infrastructure. The primary advantage of Gemma 3 is that capable multimodal AI becomes available at no licensing cost, which is meaningful for use cases where data privacy, infrastructure budget, or network access are constraints. The 2B and 7B variants run on consumer-grade hardware, while the 27B model requires more capable hardware such as an NVIDIA RTX 4090. As with most open-weight models, performance on highly specialized or low-resource languages may not match that of the largest closed commercial models. Users should verify hardware compatibility for their chosen model size before deployment.