ResearchHeinstell research
Why we train small models for big fields
Our research team on why a specialist model on the farm beats a general one in the cloud for most agronomic decisions.
General-purpose AI models are remarkable, but most farm decisions are narrow. Is this a weed or the crop? Is this cow walking differently from last week? Narrow questions suit small, specialist models, and small models can run where the questions come up.
Three things we learned
- Local data beats more data. A model calibrated on a single season of a farm’s own imagery often outperforms one trained on far more generic data.
- Latency is an agronomic variable. A decision that arrives after the sprayer has passed is worth nothing.
- Trust follows control. Farmers share more feedback, and correct more mistakes, when they know the data stays with them.
The best model for a field is the one that has seen that field.
We’ll be publishing the evaluation method behind these results later this year, including the farms and conditions it was tested in.