What Is Google Gemini? The AI Model Explained for Learners

Gemini is Google's family of large AI models, built by Google DeepMind and first announced on 6 December 2023. It was designed from the start to be multimodal, which means one model can take in text, images, audio, video and code together instead of stitching separate models together. Gemini also became the name of Google's chatbot in February 2024, when Bard was renamed.

I first wrote about Gemini the week it launched. Back then most of what anyone could say came from Google's own announcement. This version sticks to what has held up since, and tells you where to look for the details that keep changing.

Who makes Gemini and when was it released?

Gemini comes from Google DeepMind, the unit formed in 2023 when Google Brain and DeepMind merged. Google announced Gemini 1.0 on 6 December 2023 in three sizes:

• Ultra, the largest, for the most complex tasks.
• Pro, a balance of quality and cost, which went into Bard straight away.
• Nano, small enough to run on a phone. Google shipped it on the Pixel 8 Pro.

In February 2024 Google released Gemini 1.5, which could handle a context of up to one million tokens in its preview. In practice that means you could hand it a very long PDF, a large code base or an hour of video in one request. Newer generations have followed since. Model versions now change several times a year, so check Google's Gemini release notes or the Google AI for Developers documentation for the current list.

What does "multimodal" mean for Gemini?

Earlier chat models were trained mostly on text, and image understanding was bolted on later. Google trained Gemini on mixed data from the beginning. When you show it a photo of a handwritten maths problem, the same model reads the handwriting, follows the working and points out the mistake.

For our students this is useful in unglamorous ways. You can screenshot a Power BI error message, a chart, or a whiteboard from class and ask for an explanation.

How did Gemini compare with GPT-4 at launch?

Google's technical report claimed Gemini Ultra scored 90.0% on MMLU, a broad multiple-choice benchmark, and called it the first model to beat human-expert level on that test. Researchers pointed out that the 90% figure used a different prompting method (chain-of-thought with 32 samples) from the 5-shot score usually quoted for GPT-4. So the headline comparison was not like for like.

My view, then and now: benchmark tables are a starting point. Try both models on your own real tasks before deciding. I go through this in more detail in Gemini vs GPT-4.

Where can you use Gemini today?

• The Gemini app on the web, Android and iOS, for chat.
• Google Workspace, inside Gmail, Docs, Sheets and Slides, depending on your plan.
• Google AI Studio and Vertex AI, for developers who want to call the models from their own code.
• Android devices, where smaller Gemini Nano models run on the phone itself.

What are Gemini's limitations?

Gemini shares the weaknesses of every large language model. It can state wrong facts confidently. It can be thrown by ambiguous questions. It reflects biases in its training data. Google itself paused Gemini's image generation of people in February 2024 after it produced historically inaccurate images, which is a good reminder that these systems need testing in your context, not just trust.

There are practical limits too. The largest models cost more per request, and data handling depends on whether you use the consumer app or a business account, so read the terms before sending company data.

Should you learn Gemini or ChatGPT?

Learn the skills that carry across both: writing clear prompts, giving the model your own data, checking the output and knowing when not to use AI. The tools change every few months. Those habits don't.

That is the approach we take in the Generative AI training at TechNet Consultancy, where we use more than one model side by side. If you are new to these tools, start with our practical guide on using ChatGPT well.

FAQ

Is Google Gemini free?

The Gemini app has a free tier. Google also sells paid plans with access to more capable models and higher limits. Check Google's current plan pages, as names and prices change.

Is Gemini the same as Bard?

Bard was Google's earlier chatbot. Google renamed it Gemini in February 2024, so the chatbot and the model family now share a name.

What is Gemini Nano?

Gemini Nano is the smallest model in the family, designed to run directly on devices such as Android phones. It handles tasks like summarising recordings or suggesting replies without sending everything to the cloud.

Can Gemini read images and PDFs?

Yes. Gemini accepts images, documents, audio and video along with text, and can answer questions about them.

Is Gemini better than ChatGPT?

Neither wins on every task, and the lead changes with each release. Test both on the work you actually do.