Papers · Gemini
AI PAPER

Gemini: A Family of Highly Capable Multimodal Models

Introduced Google's natively multimodal model family that processes text, images, audio, and video interleaved, achieving state-of-the-art results across 30+ benchmarks.

Authors
Google DeepMind
Institution
Google DeepMind
Published
December NaN, 2023
Citations
3,500
Category
Multimodal
Impact
major
PAPER EXPLAINED

Plain-English Summary

Gemini was Google DeepMind’s entry into the frontier model race, designed from the ground up to be natively multimodal. Unlike models that bolt vision onto a language model, Gemini was trained jointly on text, images, audio, and video from the beginning. The Ultra variant achieved state-of-the-art results on 30 out of 32 benchmarks, including the first model to exceed human expert performance on the MMLU benchmark.

The model family includes three sizes: Ultra (largest, for complex tasks), Pro (balanced for scaling), and Nano (efficient for on-device deployment). This tiered approach reflected Google’s strategy of deploying AI across its entire product ecosystem.

Key Innovation

Native multimodality means Gemini processes different input types through a unified architecture rather than separate encoders bolted together. Images, video frames, audio spectrograms, and text tokens all flow through the same Transformer, enabling the model to reason across modalities naturally. This architecture enables tasks like understanding a video while reading on-screen text and processing the audio track simultaneously.

The long-context capabilities of Gemini 1.5 Pro (handling up to 1 million tokens) demonstrated that the architecture could scale to unprecedented input lengths while maintaining reasoning quality.

Impact on the Field

Gemini represented Google’s full competitive response to GPT-4 and established that the frontier model race would be a multi-player competition rather than an OpenAI monopoly. The model was integrated across Google products (Search, Workspace, Android) faster than any previous AI system, demonstrating the advantage of a vertically integrated company deploying AI at scale.

The paper’s multimodal benchmarks helped establish evaluation standards for models that process multiple input types, pushing the field toward more comprehensive capability measurement.

Models That Built on This

Gemini 1.5 Pro extended the context window to 1 million tokens and improved efficiency. Gemini 2.0 added agentic capabilities and real-time interaction. Gemma provided open-weight models derived from Gemini research. The model powers Google’s AI products including AI Overviews in Search, Gemini in Workspace, and Android’s on-device AI features.