we are preparing the receipts people need before trusting any home-model claim: power draw tokens/sec BPD/perplexity memory footprint training cost inference cost hardware config baseline comparisons claims need measurements.

Priscus model family: P1 Mini P1 P1 Max same research direction across the line: adaptive computation, specialized experts, lower footprint, local inference. mythos-class model at home.

efficiency should be designed into the model, not bolted on after training. Priscus P1 is not a bigger model squeezed down. it is a specialized expert system built to route compute deliberately from the start.

not every token deserves the same amount of compute. Priscus P1 is built around adaptive computation: spend more effort where the problem is harder, spend less where it is not. that is the core idea behind mythos-class models at home.

what would make you trust a home-runnable model claim? power draw? tokens/sec? BPD/perplexity? side-by-side evals? installer video? full hardware bill? we are deciding what to publish first.

the next model jump should not only be bigger context and bigger bills. it should be better computation. less wasted prediction. less wasted power. more useful work per watt. that is the Priscus P1 thesis.

mythos-class model at home. Priscus is building P1: a specialized expert system for adaptive computation. The goal is frontier-class utility with a fraction of the footprint. P1 Mini, P1, and P1 Max. priscus.ai