Introducing Orca 1.3

We started Oceanir with a controversial thesis: General-purpose vision models are too broad to be useful for professional geolocation. They know "everything" a little bit, but nothing deeply. To get utility, you need specialization.

So we built Orca 1.2 exclusively for Miami. It was hyper-specialized. It knew the difference between a palm tree in Coral Gables and one in Wynwood. It was accurate, fast, and completely unscalable. It was proof that if you constrain the problem space, you can achieve superhuman accuracy.

The problem with Orca 1.2 was that it didn't actually "know" what a street was. It just knew what Miami's streets looked like. When we showed it a photo of Lisbon, it tried to map it to South Beach.

To go global, we couldn't just "add data." We had to fundamentally change the architecture. We needed a model that could separate style (architecture, vegetation, light) from structure (roads, buildings, sky).

Orca 1.3 is the result of this pivot. Instead of memorizing landmarks, it learns heuristics:

• Blue street signs often imply specific regions in France.
• Yellow curbs combined with double-yellow lines are distinct to certain Commonwealth nations.
• High-contrast shadows indicate latitudinal proximity to the equator.

This is the difference between memorization and intelligence.

We didn't pick our new regions at random. We chose "adversarial" cities to stress-test the new model:

Orca 1.3 achieves a significant lead in global precision by leveraging cross-continental training data that competitors lack.

Orca 1.3 introduces six core capabilities designed for global-scale visual intelligence.