Training the next
foundation model

› How big is the opportunity?

Every company that runs a database.

The last modality without a foundation model

› Why now?

The architecture arrived in 2025

ICLR 2026. Scaling laws at ICML 2026.

› What is defensible?

Weights that generalise to unseen schemas

93% of supervised accuracy, zero labels.

› Why not just use an LLM?

Smaller models with higher accuracy create scale.

A 27B LLM gets 84%. A 22M model gets 93%. 200k+ tk/s.

› How far can this go?

Pretraining loss follows a power law

Synthetic 68.2, plus real 71.8, tuned 79.7.