GLOSSARY
Native Multimodality
One model that sees images, hears audio and writes text in the same forward pass — not a pipeline of specialists glued together.
Early multimodal products were pipelines: an OCR model, a captioner, an LLM. Native multimodality trains one network on mixed tokens — image patches, spectrograms, text — so the same weights can ground a sentence in a screenshot. Gemini, GPT-4o and similar models are sold on that unity.
The gain is cross-attention that was never bolted on: “what is wrong with this chart” does not lose the pixels in a captioning round-trip. The cost is data and evals that must cover every modality pair, and products that still silently fall back to a pipeline when the native path fails. Multimodal in the brochure is not always multimodal in the trace.