The way this actually works is hiding in a sentence from Alphabet’s FY2024 10-K: in 2024 the company “launched Gemini 2.0,” which “can generalize and seamlessly understand, operate across, and combine different types” of information. That phrase, “combine different types,” is the entire definition of a word you keep hearing: multimodal.

Under the hood, the earliest large models were unimodal, they did one kind of input. A language model read and wrote text; an image model handled pictures. Multimodal means one model takes several input types, text, images, audio, video, and reasons across them together, so it can, say, look at a photo and answer a question about it in words.

“Gemini can generalize and seamlessly understand, operate across, and combine different types of information including text, code, audio, image, and video.”— Alphabet Inc. Form 10-K (FY2024) source

The filing is precise about the sequence, and the precision matters. In 2023, the report says, Alphabet took “a significant step on our journey to make AI more helpful for everyone with the introduction of Gemini, our natively multimodal AI model.” In 2024, it “launched Gemini 2.0, our most capable model yet.” Two words in that history carry the weight: natively multimodal. Gemini is not a text model with an image add-on bolted on afterward; the company is telling investors the model was built from the start to take text, code, audio, image, and video as first-class inputs. The list of five modalities in the quoted sentence is the concrete content behind the abstract word.

The mechanism that makes this possible is representation. Each input type gets converted into the same internal numerical form, so once a picture and a sentence are both turned into the model’s shared “language,” the model can relate them. Forget the name for a second: it is one brain that learned to read several kinds of input in a common code. That is why the filing can claim the model will “generalize” and “operate across” types, the generalization is only possible because the types meet in a single representational space.

Why is this in a 10-K, and why does the scope it claims matter to a reader? Because Alphabet is telling investors that its flagship model is built around this capability, not bolting it on, and then it quantifies the reach. The report states that “all seven of our two billion-user products, Android, Chrome, Gmail, Maps, Play Store, Search, and YouTube, are using Gemini,” and that Google Cloud carries the same models to enterprise customers. The multimodal claim is therefore not a lab boast; the filing ties it to the company’s entire consumer surface. The same section reaches beyond products to science, citing DeepMind’s AlphaFold and the 2024 introduction of AlphaFold 3 to “predict the structure and interactions of all the molecules in life’s processes,” framing multimodality as one piece of a broader AI-first strategy.

There is a useful tension worth naming between what the filing claims and what a model genuinely does. “Generalize and seamlessly understand, operate across, and combine” is corporate prose, but each verb maps to a real capability. “Combine” is the literal multimodal claim, several input types entering one model. “Operate across” is the harder one: it implies the model can carry information from one modality into another, reading a chart and writing a paragraph about it, or hearing a question and pointing to the region of an image that answers it. “Generalize” is the bet that learning in a shared representation transfers, that lessons drawn from text help the model reason about images and vice versa. A 10-K is not a technical paper, but the verbs are not empty; they describe what a natively multimodal model is supposed to buy you.

Reading the section as an investor document also explains why the modality list is exhaustive rather than illustrative. “Text, code, audio, image, and video” is a deliberate enumeration: code is called out separately from text because programming is a distinct, lucrative capability; video is the most demanding modality and its inclusion signals ambition. Pair that with the claim that all seven two-billion-user products already run Gemini, and the strategic message is that multimodality is not a research line item but the substrate of the company's consumer and cloud businesses. The filing is staking shareholder expectations on the proposition that the generalist model, not a fleet of specialists, is the product.

Finally, the placement of this language inside a risk-and-strategy document, rather than a product blog, changes how literally to read it. A 10-K is reviewed by lawyers and auditors; the claim that Gemini is “natively multimodal” and runs across all seven flagship products is a representation to the market, not marketing copy that can be quietly revised. That is exactly why it makes such a clean teaching example: the definition of multimodality, one model that can “combine different types of information including text, code, audio, image, and video”, is stated plainly, dated to the fiscal year, and tied to named products and a named model version. For a reader trying to separate AI hype from AI substance, a company's own annual filing is among the most disciplined places the claim ever gets written down.

As a plain-language marker for early 2025: “multimodal” is not jargon for jargon’s sake, it is a model that stopped being a specialist and became a generalist across input types, and a company’s own annual filing is a clean place to see the claim stated plainly, dated, and tied to named products. When the 10-K leads its AI narrative with “understand, operate across, and combine different types,” Alphabet is staking its model strategy on multimodality being the direction the field is heading, and putting its two-billion-user products behind the bet. Filing data and the evidence index via SEC filings.