Google Gemini — AI Glossary

Reference details

Related terms

• Google Vertex AI
• Multimodal Architecture
• Long-Context Model

How it works

1. The family is natively multimodal: text, images, audio and video are processed by one model trained on them together rather than by a language model with an encoder attached, which is what allows interleaved inputs in a single request.
2. Size tiers span from on-device to frontier, sharing an API shape so that a request written against one runs against another by changing a model identifier.
3. Very long context is the family's distinguishing capability, with limits an order of magnitude above the usual range — which changes what is worth retrieving versus simply including.
4. Tool use and streaming are first-class in the later generations, so the model can call functions and produce output incrementally within one session.
5. Access routes differ in what they expose: the developer studio, the cloud platform and on-device integration do not all offer the same tiers, limits or data-handling terms.
6. Dated model identifiers are what pin behaviour, since the family and tier name a lineage and a size rather than the artefact that will answer a request.

Example: What a million-token context actually costs per request

Suppose input is billed at 1.25 units per million tokens. Filling a one-million-token context costs 1,000,000 x 1.25 / 1,000,000 = 1.25 units for the context alone, before the model has produced anything. Across 10,000 requests that is 10,000 x 1.25 = 12,500 units. Retrieving a 32,000-token slice instead costs 32,000 x 1.25 / 1,000,000 = 0.04 units per request, or 10,000 x 0.04 = 400 units — a factor of 12,500 / 400 = 31. So the long context is not a free capability but a priced one, and the comparison against retrieval is a volume question in the same shape as any other: at 10,000 requests a 12,100-unit difference buys a great deal of retrieval engineering, and at 100 requests it does not. The capability earns its cost specifically where the whole context is genuinely needed — a document whose parts interact, a codebase where the relevant file is not predictable — and loses it where a retriever would have found the same 32,000 tokens.

Commonly confused with

Natively multimodal and multimodal-by-adapter describe different training histories, and the difference shows up in cross-modal tasks rather than in single-modality ones. A model with a vision encoder bolted onto a language model learns the alignment after the fact, so the two modalities meet at a projection layer. A natively multimodal model is trained on the modalities together, so representations are shared from the start — which matters for tasks requiring reasoning across modalities in one step, such as relating a spoken instruction to a region of a video. For single-image question answering the distinction is much less visible, which is why architecture claims should be read against the task that would show them.

When to use it

Reach for it when:

• Interleaved multimodal input in a single request — text with images, audio or video together.
• Genuinely long-context work, where the parts of a large document interact and retrieval would have to fetch most of it anyway.
• On-device requirements, where the smallest tier is the reason to be in this family at all.
• With a dated identifier pinned, and with the access route chosen deliberately since tiers and terms differ by route.

Reach for something else when:

• Filling the context because it is available, which is a priced capability rather than a free one.
• As a capability claim from the family name, since tier and generation determine behaviour and the benchmark pages hold the measured figures.
• Assuming parity across access routes, which differ in available tiers, limits and data handling.
• Configured at family or tier level in production, which leaves the actual artefact unpinned.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Google Gemini — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path