Why trust Greenbot
Source: Google for Developers
Google launched Gemma 3, a new artificial intelligence (AI) model achieving 98% of DeepSeek R1‘s accuracy while using just one NVIDIA H100 GPU—compared to R1’s estimated 32 GPUs.
This leap in efficiency could make high-performance AI far more accessible.
In Chatbot Arena Elo ratings, Gemma 3 scored 1338 compared to DeepSeek R1’s 1363. While R1 maintains a slight lead, Google’s achievement focuses on delivering similar results with far less computing power.
Source: Google
“Gemma 3 delivers state-of-the-art performance for its size, outperforming Llama-405B, DeepSeek-V3, and o3-mini in preliminary human preference evaluations,” Google stated in its developer blog. “This helps you to create engaging user experiences that can fit on a single GPU or TPU host.”
The model features four sizes (1B, 4B, 12B, and 27B parameters) built to run on devices from phones to workstations. Even its largest version is dramatically smaller than DeepSeek R1’s 671 billion parameters.
Gemma 3 provides “out-of-the-box” support for 35 languages with pre-trained capabilities for 105 more. This represents a dramatic expansion from Gemma 2, which primarily supported English.
Source: Google
Beyond language support, Gemma 3 brings multimodal capabilities. This means the AI can understand and work with words, pictures, and videos all at once, rather than just processing text. Its expanded 128K token context window processes approximately 300 pages of text or 30 high-resolution images at once.
These efficiency improvements come from model distillation and training techniques like Reinforcement Learning from Human Feedback (RLHF) and Reinforcement Learning from Machine Feedback (RLMF). Full technical details are available in Google’s 26-page technical report on the Gemma website.
Google paired the release with ShieldGemma 2, a specialized 4B parameter model for content safety. The company’s technical report notes, “We find that Gemma 3 models memorize long-form text at a much lower rate than prior models,” suggesting better data privacy than earlier versions.
The push for AI efficiency has grown stronger with Microsoft’s Phi-4 and Mistral Small 3 joining the competition. These smaller models target businesses that need AI capabilities without the massive computing resources of data centers. Running advanced AI on a single GPU dramatically reduces the entry barriers for smaller organizations.
Developers can access Gemma 3 through Google AI Studio, Hugging Face, and Kaggle. Academic researchers can apply for Google’s program, which offers $10,000 in cloud credits for Gemma 3 research.
While Google calls Gemma “open,” some debate exists around its licensing terms restricting certain uses. Still, the model represents progress in making advanced AI work with reasonable computing resources.
