Google introduces Gemini 1.5 Flash at I/O 2024

What happened

On 14 May 2024, at its I/O event, Google introduced Gemini 1.5 Flash in a post by Demis Hassabis.^[1] Google described the model as lighter-weight than Gemini 1.5 Pro and as the fastest Gemini model served in its API.^[1]

A smaller model built from a larger one

Two design choices define the release. The first is distillation: Flash was distilled from Gemini 1.5 Pro, meaning the smaller model was trained to follow the behaviour of the larger one rather than being designed from scratch as an independent system.^[1] Distillation is a common way to obtain a model that is cheaper and quicker to run while keeping much of the larger model's behaviour; what it gives up, in general, is the ability to match the teacher everywhere, since the student has less capacity to work with.

The second is the context window — the span of text a model can take into account at once. Flash was offered with a 1-million-token context window.^[1] A token is the unit in which models read text, roughly a short word or a word fragment; context length matters because material outside the window cannot influence the answer. On the same day, Google made a 2-million-token context window for 1.5 Pro available through a waitlist, so the larger window was an option users had to request rather than the default.^[1]

The other announcements

Flash was not the only item. Google previewed Project Astra and described it as its vision for the future of AI assistants.^[1] It also announced Gemma 2, saying the model would have a new architecture designed for breakthrough performance and efficiency and would come in new sizes, and it announced PaliGemma.^[1] Taken together, the announcements covered four different things: a fast hosted model in the API, a much longer context option on the larger model, open models under the Gemma name, and a research preview of an assistant prototype.^[1]

Why it matters

Naming a model Flash and shipping it alongside a larger sibling makes the trade-off explicit: users pick between a model that is faster and lighter and one that is larger, and Google's own description framed Flash by comparison with 1.5 Pro.^[1] Context length was presented as a separate axis of progress, since the waitlist for a 2-million-token window applied to the Pro model rather than to Flash.^[1] A single event that also carried an assistant preview and new open models shows that a model launch now sits inside a wider product line rather than standing on its own.^[1]