Aiodic

Back to the table

59 Model family Ge Gemini Model Families

Gemini

Google's model family.

What it is

Gemini is Google's family of AI models, and also the name of the chatbot app that uses them. It is made by Google DeepMind, the research lab Google formed in April 2023 by merging its two AI teams, Google Brain and DeepMind, under DeepMind's co-founder Demis Hassabis. Like GPT from OpenAI and Claude from Anthropic, Gemini is a language model at heart: a system trained on enormous amounts of text and other material so that it can read, write, summarise, answer questions, and help with code. What Google has emphasised from the start is that Gemini was built to handle pictures, sound, and video as well as words.

"Gemini" is used for three different things, which can be confusing at first:

From Bard to Gemini

The story starts with a different name. On 6 February 2023, a little over two months after OpenAI's ChatGPT became a sensation, Google's chief executive Sundar Pichai announced Bard, an "experimental conversational AI service" based on an earlier Google model called LaMDA. It went first to a small group of trusted testers. On 21 March 2023, Google opened Bard to the public in the US and the UK, with a waiting list, and its own launch post warned that it could give inaccurate information while sounding confident.

On 6 December 2023, Google and Google DeepMind announced Gemini 1.0 as their new model family, and Bard switched to a tuned version of it that day. Two months later, on 8 February 2024, Google retired the Bard name entirely. The chatbot became Gemini, a paid tier called Gemini Advanced gave access to the largest model, and a Gemini app arrived on Android phones.

Sizes and generations

The first generation came in three sizes, which Google described this way: Ultra, the largest, for highly complex tasks; Pro, for a wide range of tasks; and Nano, the smallest, built to run on a phone itself rather than on Google's computers. The Pixel 8 Pro was the first phone Google engineered to run Nano, using it for things like summarising voice recordings.

Since then the names have kept moving, which is normal for this field:

Ultra faded as a model name after the first generation, while Pro, Flash, and a cheaper Flash-Lite became the usual tiers. Model names, versions, and which one the app uses by default change every few months, so check Google's current model list rather than relying on any article, including this one.

What "natively multimodal" means

"Multimodal" means able to work with more than one kind of input, such as text, images, audio, and video. For a while, the usual way to build such a system was to bolt separate parts together: one program to describe an image in words, another to transcribe speech, and a language model to read the result. Each hand-off loses something, much as a story loses detail when it passes through several people.

Google's announcement in December 2023 said Gemini was different: "natively multimodal", trained from the start on different kinds of material together, rather than stitching separate components together afterwards. In everyday terms, the model looks at your photo of a handwritten recipe directly, instead of relying on a separate step to turn the photo into text first. That's why you can show Gemini a chart, a screenshot, a page of notes, or a short video and ask questions about it in the same conversation.

A sibling: Gemma

Google also publishes Gemma, a family of smaller "open" models, first released on 21 February 2024 and described by Google as built from the same research and technology as Gemini. The difference is who runs them. Gemini runs on Google's computers and you reach it through Google's apps or its API, a way for software to talk to other software. Gemma models can be downloaded and run on your own computer or server, like Llama from Meta.

Why it matters

It is where a huge number of people meet AI. Google puts Gemini inside products people already use every day, so many people use it without ever opening a chatbot. At the time of writing, AI answers appear at the top of many Google searches, and a separate AI Mode in Search, opened to all US users in May 2025, lets you have a back-and-forth conversation instead of getting a list of links. In November 2025 Google said the Gemini app had passed 650 million monthly users. Those are Google's own figures, and companies count "users" in different ways, but the scale is plainly very large.

It reaches into your phone and browser. In March 2025 Google announced that Gemini would replace the older Google Assistant on most Android phones, and it later said the change would finish in 2026. Since September 2025, Gemini has also been built into the Chrome browser for desktop users in the US, where it can explain the page you are reading. Details differ by country, device, and account type, so check what is on yours.

Multimodality changes what you can ask. For a non-technical person, the most useful thing about Gemini is often not writing, but seeing. You can point your camera at a plant, a confusing parking sign, or an error message on a screen and ask what it means. Business owners can hand it a photo of a whiteboard, a scanned invoice, or a product label. Once you know the AI can read images, a lot of small daily jobs look different.

It plugs into the rest of your Google account. If you use Gmail, Google Drive, Docs, or Calendar, Gemini can work with that material when you allow it, for example summarising a long email thread or drafting a document from your notes. That is convenient, and it is also a reason to think carefully about privacy, which we come back to below.

It keeps the competition honest. Google, OpenAI, Anthropic, Meta, and others release new models every few months, and each claims leadership on some tests. No single company stays ahead on everything for long, and claims of being "the best" come from the companies themselves, measured on benchmarks they choose. For users, that competition has meant faster improvements and falling prices. The practical lesson is to try more than one model for the jobs you care about.

Real-world examples

The telescope answer that cost billions (February 2023)

Bard's very first public appearance included a mistake. Google posted a short promotional demo on Twitter in which someone asked what new discoveries from the James Webb Space Telescope they could tell a nine-year-old about. One of Bard's answers said the telescope had taken the very first pictures of a planet outside our solar system. That was wrong: the first such image was taken in 2004 by the European Southern Observatory's Very Large Telescope in Chile, long before Webb launched.

Reuters spotted the error on 8 February 2023, the same day Google held an AI event in Paris. Alphabet's shares fell by roughly 8 percent that day, wiping about 100 billion dollars off its market value. A Google spokesperson said the mistake highlighted "the importance of a rigorous testing process", which the company was starting that week with its trusted testers. The episode is explained as a lesson about made-up facts in Hallucination. For the Gemini story, it matters for another reason: it showed how much pressure Google was under to answer ChatGPT quickly, and it set the tone for how closely every later launch would be examined.

The demo video that wasn't live (December 2023)

When Google launched Gemini on 6 December 2023, it published a video called "Hands-on with Gemini". It appeared to show someone talking to Gemini while drawing a duck, playing rock, paper, scissors with their hands, and hiding a ball under cups, with Gemini commenting smoothly in real time. It gathered around a million views within days.

On 7 December, TechCrunch reported that the interactions had not happened the way the video suggested, and Google confirmed to Bloomberg that the demo had been made using still image frames from the footage and typed text prompts. The video's YouTube description did say that response times had been reduced and Gemini's answers shortened, but the video itself didn't make that clear. Google said the outputs shown were real.

This is a useful example precisely because of what Gemini was being sold on. Multimodality is real, and the model could answer questions about those images. But a model that can describe a still photo when you type a careful question is not the same as one that watches you live and chats back. When you see any AI demo, from any company, ask how it was made before you judge what the product can do.

Other moments covered elsewhere on this table

Several important Gemini moments have their own full stories in other elements. Google's launch claim that Gemini Ultra scored 90 percent on the MMLU exam, and the fine print about how that score was reached, is in Self-consistency. The February 2024 pause of Gemini's ability to generate pictures of people, after it produced historically inaccurate images, is in Bias. Gemini's Deep Research feature is in Research, the thinking budgets of Gemini 2.5 are in Extended Thinking, custom Gems are in Chatbots, and the notebook tool that answers only from your own sources is in Grounding.

An everyday case: the shoebox of receipts

Picture Marcus, who runs a small painting and decorating business and keeps his receipts in a shoebox until his accountant chases him. One evening he photographs a dozen receipts on the kitchen table and uploads them to Gemini with a request: list the supplier, date, total, and tax on each one, as a table. Most come back right. Two faded ones are misread, a 7 comes out as a 1, and one receipt is listed twice because it appeared in two photos. Marcus catches these only because he checks each line against the paper. He then uses the table as a starting point for his bookkeeping, not as the final record, and he crosses out his card number on each receipt before taking the next batch of photos.

Try it yourself

This exercise uses Gemini's ability to read images. It takes about fifteen minutes and works with the free version of Gemini at the time of writing. Menus and labels change often, so don't worry if yours look a little different.

A privacy note first. Don't upload anything with card numbers, account numbers, ID documents, medical details, or other people's personal information. Google's privacy page for Gemini says that a sample of chats is read by trained human reviewers, that reviewed chats can be kept for up to three years even if you delete your activity, and it asks you not to enter anything confidential you wouldn't want a reviewer to see. Use a receipt for something ordinary, or a page from a printed leaflet or book, and cover or crop out anything personal.

  1. Open Gemini. Go to gemini.google.com in a browser, or open the Gemini app on your phone, and sign in with a Google account.
  2. Consider a temporary chat. At the time of writing, Gemini offers a temporary chat option that isn't saved to your history. If you see it, use it for this exercise.
  3. Take a clear photo. Lay a receipt or a printed page flat in good light and photograph it straight on. Blurry or angled photos cause most mistakes.
  4. Upload it and ask for the facts. Use the plus or attach button to add the photo, then paste this prompt: > Read this image carefully. List every item and price you can see, then the total. If any part is hard to read, say which part and give your best guess marked as a guess. Don't fill in anything you can't actually see.
  5. Check it against the paper. Go line by line. Did it misread any numbers? Did it invent an item? Did it say when it was unsure?
  6. Ask a question that needs understanding, not just reading. For a receipt, try: > Does the total match the sum of the items? Show your working. If there is tax or a discount, explain how it was applied. For a printed page, try: > Explain the main point of this page in three plain sentences for someone who has never read about this topic. Then list any words on the page a beginner might not know, with a short definition of each.
  7. Test its honesty. Ask something the image can't answer, such as "What time did I make this purchase?" when no time is printed, or "What does the next page say?". A good answer admits it can't know. If it makes something up, you have just seen hallucination at work.
  8. Try one more type of image. Photograph a chart from a newspaper, a screenshot of a confusing settings menu, or a label on a household product, and ask what it means. Notice where the model is strong (reading clear printed text, explaining ideas) and where it slips (small print, handwriting, anything partly hidden).
  9. Delete what you don't need. When you're done, delete the chat, and look at your Gemini activity settings to decide what you want kept. The settings and how long things are kept can change, so read Google's current privacy page yourself.

Common mistakes and misunderstandings

Thinking "Gemini" means one thing. It is a family of models, an app, and a set of features inside other Google products, and they aren't always running the same model. The free app, the paid plans, Search, Gmail, and the developer API can each use a different Gemini version at any moment. If a friend says "Gemini couldn't do that", it's worth asking which Gemini, and when.

Trusting a reading of an image because it looks confident. Reading a photo is not the same as being sure. Faded print, handwriting, glare, and cut-off edges all cause errors, and the answer will often sound just as confident either way. For anything with money, dates, or dosages in it, check against the original.

Uploading private documents without thinking. A photo can carry far more than you meant to share: a card number at the bottom of a receipt, an address on an envelope in the background, a child's name on a school letter. At the time of writing, Google says some chats are reviewed by people and that uploaded images are treated like any other prompt. Crop first, and read the current privacy terms.

Assuming free and paid are the same model. At the time of writing, Google offers a free version of Gemini and several paid Google AI plans with higher limits and access to its most capable models and features. What each plan includes changes, so check Google's current plan page before deciding a feature doesn't exist or paying for one you won't use.

Believing launch-day claims as settled fact. Every AI company, Google included, launches models with charts showing it ahead on chosen tests, and every one of them has had a demo or a number questioned afterwards. Treat benchmark scores and demo videos as the maker's claims, and judge a model by how it does on your own tasks.

Confusing Gemini with Gemma. Gemini runs on Google's servers and is reached through Google's apps and API. Gemma is a separate family of smaller open models you can download and run yourself. They share research, not a login.

Expecting the answer to be current. Gemini can search the web in many situations, but not always, and a model's built-in knowledge stops at the date its training ended. When facts matter, ask it to show its sources, and open them. See Web Search and Grounding.

Related elements