Agentic visual reasoning with code execution
Google's largest and most capable AI model. Built from the ground up to be multimodal, Gemini can generalize and seamlessly understand, operate across and combine different types of information, including text, images, audio, video and code.
Gemini's multimodal training enables it to recognize and understand text, images, audio simultaneously, making it especially skilled at explaining reasoning in complex subjects. Its ability to think carefully before answering difficult questions leads to significant improvements over first-impression responses.
Gemini's three optimized sizes (Ultra for complex tasks, Pro for scaling, Nano for on-device) enable flexible deployment across infrastructure tiers. This scalability allows enterprises and developers to run the same model family efficiently on cloud servers or edge devices without retraining.
Gemini's native multimodal architecture enables seamless understanding of text, images, audio, and video together. Its sophisticated reasoning capabilities extract knowledge from hundreds of thousands of documents, making it ideal for uncovering insights across mixed-media datasets in fields like science, finance, and research.
Gemini Ultra achieves state-of-the-art performance on MMLU (90.0%, first to exceed human experts) and MMMU benchmarks, demonstrating advanced reasoning across 57 subjects including math, physics, and ethics. It generates high-quality code in Python, Java, C++, and Go, making it suitable for technical problem-solving.
Gemini was designed from the ground up as natively multimodal with pre-training across different modalities, eliminating the need to stitch together separate components. Developers can leverage Gemini Ultra, Pro, or Nano variants to build applications with state-of-the-art multimodal reasoning without architectural workarounds.