Google DeepMind introduces Gemini

Background

Large language models (LLMs) are neural networks trained on very large collections of text to predict the next word. They became the core of many modern AI assistants. Before Gemini, many multimodal systems were assembled from separate parts: one component for text, another for images, and others for speech. Each hand-off between components could lose information.

Google had its own chatbot, Bard, and had been building language models for some time. Gemini was introduced as a new model family for that work.^[1]

What happened

On 6 December 2023 Google announced Gemini 1.0 in a post by Sundar Pichai and Demis Hassabis. The company described it as its most capable and general model yet.^[1] Gemini 1.0 came in three sizes: Ultra for highly complex tasks, Pro for scaling across a wide range of tasks, and Nano for tasks that run on a device.^[1]

Key details

The central design choice was that Gemini was built to be natively multimodal. According to Google, it was pre-trained from the start on text, images, audio and video, rather than assembled from separately trained components.^[1] In plain terms, the model learned to relate these kinds of information during training, instead of relying on a translation step between parts.

Google reported that Gemini Ultra exceeded previous state-of-the-art results on 30 of 32 widely used academic benchmarks. On MMLU, a multiple-choice test covering many school and professional subjects, Ultra scored 90.0%. Google said this made it the first model to outperform human experts on that benchmark.^[1] Benchmark scores measure performance on fixed tests, so they show strength on those tests rather than on every real-world task.

Gemini was trained on Google's TPU v4 and v5e chips, and the same announcement introduced Cloud TPU v5p.^[1] A TPU (tensor processing unit) is a chip Google designed for the matrix arithmetic that neural networks depend on.

Access rolled out in stages. Bard began using a fine-tuned version of Gemini Pro in English in more than 170 countries and territories. Fine-tuning means continuing training on a narrower set of examples to suit a particular purpose. The Pixel 8 Pro became the first smartphone built to run Gemini Nano.^[1]

From 13 December 2023, developers and enterprise customers could access Gemini Pro through the Gemini API in Google AI Studio or Vertex AI. Ultra was first offered to select customers, developers and partners.^[1]

Impact

The release put native multimodality at the centre of a major lab's flagship model family. Its tiered sizes also set out a pattern of one design offered at several scales, from large data-centre workloads down to phones. The staged rollout showed a common path for frontier models: first a chatbot, then devices, then developer access, with the largest version reserved for selected partners at first.^[1]

The figures in this article are Google's own reports at launch, not independent evaluations. Later Gemini releases built on this first generation, and readers comparing claims over time should keep that distinction in mind.