Definition
Gemini is Google DeepMind's family of models for language and multimodal tasks. Google also uses Gemini as the name of its assistant application. A report about a Gemini model and a report about the Gemini app can describe different systems, with different tools and operating conditions.
A family member's supported inputs and outputs depend on its release. Multimodal does not mean every Gemini endpoint can accept or generate every kind of media. The API catalog specifies those capabilities separately.
Origin and attribution
Google introduced the Gemini family on December 6, 2023. The announcement, presented by Sundar Pichai and Demis Hassabis, credited Google DeepMind and described the original Ultra, Pro, and Nano members.
On February 8, 2024, Sissie Hsiao announced that Google's Bard application would become Gemini. The model-family introduction and the application rename are separate events. Bard is therefore a former application name, not an alternate name for every Gemini model.
Family members and scope
Gemini Pro, Flash, and Flash-Lite identify different model tiers within the family. A version such as Gemini 3.1 Pro or Gemini 3.8 Flash adds a generation and release distinction. Live and speech variants can have separate interfaces and capabilities. Compare tiers within the stated generation and task; their names establish no permanent ranking across releases.
Google's catalog distinguishes stable and preview availability. A production evaluation needs the full model identifier and any reasoning settings, alongside the input and output modalities used. Availability and retirement policies can change independently of the family definition.
Gemma is Google's separate family of models with downloadable weights, developed using research and technology also used for Gemini. A Gemma release does not establish downloadable access to a similarly named Gemini endpoint.
Operational significance
Preserve the model and application configuration when comparing results. A Gemini app session with tools, saved instructions, or connected data can provide capabilities that a bare API request lacks. A larger advertised context window still needs an evaluation of whether the system finds and uses the relevant evidence inside that context.
Check generated claims and test proposed actions through the factory's own controls. The vendor family name establishes neither a correctness threshold nor permission to execute a tool.
Distinguish it from nearby terms
- A foundation model describes a broad reusable model class. Gemini names one family within that class.
- A multimodal model handles multiple data types; each Gemini release specifies which combinations it supports.
- A chatbot is the surrounding conversational application. The Gemini app can add behavior beyond one model call.
Check your understanding
A team repeats a Gemini app demonstration using a Flash API model and gets a different result. Which differences in model version, modalities, tools, and supplied context should it inspect, and what would a controlled comparison measure?