Omnimodel

An omnimodel isn't a model that does just one thing well. Instead, it packs several capabilities — seeing, understanding, predicting — into a single model. Think of a Swiss Army knife: instead of carrying a separate knife, scissors, and screwdriver, you fold multiple tools into one handle.

For example, teams building robots or self-driving cars used to need separate models: one to recognize objects in front of them, another to predict what happens next, and another to decide how to move. An omnimodel bundles these different roles into a single model family, so they can be picked individually or linked together as needed.

When multiple capabilities share one model like this, they all end up understanding the world in the same way. Since the part that recognizes objects and the part that predicts what happens next aren't running on different logic, the results can come out more consistent.

The article describes NVIDIA's Cosmos 3 as an open physical AI foundation omnimodel that "combines vision reasoning, world generation, and action prediction into a single model family." It's worth noting that here, "omnimodel" isn't the name of a specific product — it's a general term meaning that multiple capabilities have been integrated into one model.