FUND THE
NEXT MODEL.

The XERJ engine is open source (Apache-2.0). The XERJ Query-1 model is not open yet; it was trained mostly on rented GPUs paid for with donations, plus a few machines of our own. Every donation buys rented GPU hours for the next training program. Nothing else: no salaries, no marketing, no cloud bills for the website. We have not yet published a ledger of donations and spending.

XERJ Query-1 is a neurosymbolic model that compiles a question into a short program the engine runs over facts you have written, and shows the steps for lookups and comparisons. It was trained in FP16 on GPUs rented by the hour on Clore.ai, chosen by measured throughput per dollar, plus a few machines of our own. So far that cost thousands of dollars, not millions, paid for by donations; we have not published the exact amount yet.

The next training program is our estimate, not a measured number: about 100,000 H100 GPU-hours. It buys larger training worlds, robustness to fresh question wordings, sums and thresholds in the program vocabulary, fewer declined questions, and re-scoring on sealed evaluation sets. We publish our scores, including the ones we withdraw. The run records, evaluation data and scripts behind them are not public yet.

To everyone who has donated so far: thank you. To everyone who has donated so far: thank you. You paid for the GPU hours behind the first model. Its numbers are published in the announcement; the runs that produced them are not public yet.