Gemini Flash

Google develops a family of computer programs called Gemini. These programs answer questions, write texts, summarize documents, and analyze images. Within this family there are various sizes, similar to engines in a car lineup. Gemini Flash is the small, economical variant. It responds very quickly and costs only a fraction per request compared to its larger siblings. In return, it is less reliable than the large model called Gemini Pro on very difficult tasks, such as long math or programming problems.

Why Google builds an economy version

Every answer from such a program costs computing time in a data center. With millions of requests per day, this adds up to considerable sums. A company that automatically sorts customer inquiries doesn’t need a top-performance program for that. It needs a program that handles the task well enough while costing almost nothing.

On top of that comes the wait time, known in technical jargon as latency. If you’re typing in an app and waiting three seconds for a reaction, that feels sluggish. Flash often delivers the first words after mere fractions of a second. For autocomplete, chat translations, or voice assistants, this is crucial.

A common misconception is that a cheap model is simply a bad model. In fact, Gemini Flash beats the expensive top models from two years ago on many standard tasks. So the question isn’t which model is the best, but which one is sufficient for a particular task. This decision is called model selection, and it has since become an important cost factor in software projects.

What distinguishes Flash from Pro

Such programs consist of billions of adjustable numerical values, the parameters. They are automatically adjusted during training until the program delivers useful answers. More parameters usually mean more knowledge and better reasoning, but also more computational effort per answer. Flash has significantly fewer parameters than Pro. Google does not disclose the exact figures, which is common practice for commercial models.

An important trick is called distillation. In this process, the small model doesn’t just learn from texts but directly from the large model. The large model answers tasks, and the small one practices producing the same answers. Like a student looking over the shoulder of an experienced teacher instead of figuring everything out alone. This way, Flash achieves most of Pro’s quality at a fraction of the size.

Despite its small size, Flash can process a huge amount of text at once. For the newer versions, Google states a window of around one million text tokens, which corresponds to roughly a thousand book pages. Furthermore, Flash understands not just text but also images, audio, and video. Such models are called multimodal.

Flash in apps and in quarterly earnings

Anyone using the Gemini app or the AI summaries in Google Search often gets their answers from a Flash variant. For simple queries, Google deliberately switches on the cheap model, and only uses the large one when needed. Flash also works in the background in Google Docs, Gmail, and on Android phones.

For developers, Flash can be accessed via a programming interface, known as an API. Billing is based on the amount of text processed, often in the range of a few cents per million text tokens. It is precisely these prices that are at the center of the competition with OpenAI and Anthropic, who offer their own economy models.

In business news, Flash models therefore regularly come up in connection with two topics. First, in rounds of price cuts, when a provider reduces the cost per request. Second, in the question of whether the billions in investments in data centers pay off. Cheap models increase usage but lower the revenue per request.