Priscus — AI Research
Frontier performance.
Fraction of the footprint.
Priscus P1 is a specialized expert system built for adaptive computation — smaller, cheaper to train, and measurably more efficient. By architecture, not compression.
Priscus P1 — Headline Result
Measurably better, on less.
Measured against the leading open-source model in its size class, trained on an identical corpus. Full benchmark results under NDA.
Parameters
−26%
Same training regime, smaller model. Lower deployment cost and memory footprint at competitive quality.
Held-Out Loss
−92.7%
Cross-domain validation loss vs. the leading open-source baseline in matched training conditions. Full methodology, including vocabulary-normalized results, available under NDA.
Overfit Gap Growth
0
When training loss collapses to memorization levels, validation loss continues to decrease rather than diverge — the opposite of standard transformer behavior. Full training curves available under NDA.
↳ Full benchmark suite available to qualified investors under NDA.
What makes P1 different
Three ideas, working together, the only way we can see the next order of magnitude.
Heterogeneous Experts
Specialization earns compute.
Not every token deserves the same amount of thought. Priscus P1 routes computation across a population of specialized experts, so capacity flows where it's needed and coasts where it isn't.
Adaptive Computation
Depth, not just width.
Traditional scaling buys intelligence by making models wider. P1 buys it by letting the model decide when to think longer — graded difficulty, graded effort, per token.
Efficient by Design
Frontier loss curves. Fraction of the budget.
Our architecture achieves its results with materially fewer parameters and less training compute than comparable models. The savings compound at inference time.
Research Philosophy
We are building what transformers should have been.
We treat efficiency as a first-class design objective, not a compression step tacked onto a bigger model. Our research focuses on architectures where compute is spent deliberately, not uniformly — where a model can allocate its attention the way a human expert would.
Priscus P1 is the flagship of our first model family. A smaller variant (P1-mini) for on-device inference and a larger variant (P1-XL) for long-horizon reasoning round out the lineup — each targeting a different deployment regime. We publish what we can, protect what we must, and favor the harder, better problem.
Peer reviewed work, forthcoming
Paper · Benchmarks · Demo — Q3 2026
Team
Founder-led. Not rushing to hire.
Priscus is one person right now — focused on proving out the architecture. The next hires will be top-tier researchers and engineers who treat foundation-model efficiency as an engineering problem and a research problem at the same time. No C-suite games. No “AI strategist” roles.
The kinds of people we'd want
- ◇ResearchArchitecture, training dynamics, evaluation — the kind of person who reads papers and runs ablations in the same afternoon.
- ◇SystemsLow-level runtime, tiered memory, kernel work. C and CUDA by choice, not by obligation.
- ◇TrainingData curriculum, scaling, the long-horizon work of actually training a foundation model well.
No public postings. Quiet conversations only.
hello@priscus.ai →Investors → We're raising. Technical diligence welcomed.
Early Access
Early access to Priscus P1.
We're onboarding a small cohort of researchers, developers, and technical teams ahead of public release.
Expect the first invites in Q3 2026. Priority goes to teams doing substantive work at the frontier — whether that's research, product, or deployment.