Wait, am I misunderstanding what omnimodal actually means?Solved

You are half right, but you are confusing multimodal with true omnimodal. 

What you described (uploading an image and typing a prompt) is standard multimodal. But a lot of those older models cheat. They use a separate vision encoder or a speech to text tool (like Whisper) to translate your image or voice into a text format first, and then the core LLM processes that text. 

True omnimodal (like GPT 4o or the newer Qwen Omni stuff) means it natively processes AND generates any of those formats in the exact same neural network. There is no translation step. It does not turn your voice into text to understand you; it processes the raw audio waveform directly. It is an Any to Any architecture. Audio and video go in, audio and video come directly out. No middlemen. 

Exactly what @jayceon said. The reason it is a massive revolution is not just a technicality; it completely changes how the AI behaves. 

When an older multimodal model translates your voice to text, it immediately loses your tone of voice, your sarcasm, your breathing, and the background noise. It compresses reality into a flat string of words. 

Because true omnimodal models process the raw audio or video directly in their core brain, they actually hear the sarcasm. They hear you laugh. And because they do not have to wait for three different translation pipelines to finish, the latency drops from 4 seconds down to 300 milliseconds. That is what makes real time, human like voice agents finally possible.