Google releases Gemma 3, a low-cost, high-performance multimodal large model. A single H100 can run the most powerful open source model

PANews reported on March 13 that Google CEO Sundar Pichai announced that Google DeepMind officially released the latest open source multimodal large model Gemma 3, which focuses on low cost and high performance, and supports text, visual reasoning and function calls. Gemma 3 has four parameter scales: 1B, 4B, 12B and 27B. The largest 27B model only needs one H100 GPU to run efficiently, while similar models require at least 10 times the computing power to achieve similar performance.

Gemma 3 supports 140 languages and provides a 128K context window, significantly improving the ability to process long texts. According to the LMSYS ChatbotArena test, its 27B model is second only to DeepSeek R1-671B, and higher than well-known open source models such as o3-mini and Llama3-405B. In addition, Google simultaneously launched ShieldGemma 2, which focuses on image security detection and helps developers filter sensitive content. Gemma 3 is now available for download and can be used through platforms such as Google AI Studio, Hugging Face, and Vertex AI.