Kimi K3 won. And then something better happened.
We threw the same question at 7 models from 5 vendors: what's risky about this French address?
K3 came in hot. From nothing but a coordinate, it nailed Nice's seismic zone. The 1997 asbestos cut-off. The Cantal's radon category. Best property knowledge we've ever measured.
Then we handed it Secrets.RealEstate's property knowledge graph.
It got better anyway. 🚀
Every single model did. And here's the part that gave us chills — the smarter the AI, the smaller the gain, in a perfectly straight line. But it never hits zero. Not at 2.8 trillion parameters. Not ever. Because no model on earth invents what a flat actually sold for, or what it actually burns each winter.
Then the plot twist: grounding made it 4× faster. Eight minutes of guessing became two minutes of knowing. And K3 does it at a third the price of the frontier model it just beat.
Bigger models won't make good data obsolete.
They make it obvious what good data is worth.
