Measured
intelligence.
A family of language models trained on a single scaling curve — where every result predicts the next. Three models, reproduced within 1%, for $180 of compute.
Four points. One line.
Each model is a controlled measurement on the same data recipe. Same method, rising scale — so every point predicts the next one.
| Model | Run 1 | Run 2 (from scratch) | Δ |
|---|---|---|---|
| Seed · 45M | 3.6616 | 3.6072 | −1.5% |
| Sprout · 150M | 3.0679 | 3.0603 | −0.2% |
| Sapling · 500M | 2.7149 | 2.7116 | −0.1% |
Final validation loss. The family retrained on independent hardware landed within 1% every time — the signature of a repeatable recipe, not a lucky run.
Watch the smallest one think.
Real, unedited output from the 45-million-parameter model — running on a laptop CPU, no GPU. It's tiny, and it shows: grammar works before knowledge does.
The alphabet and simple addition emerge cleanly; ask it a fact and it improvises. That gap is exactly what more scale closes — measured on the curve above.
The same question, growing up.
Prompt: "To bake bread, you first need" — asked to each model as it grew. Watch coherence arrive before, then with, correctness.
Built to grow up.
An architecture designed for staged development — and for decoupled memory, planning, and reasoning as it scales. Here's what's measured, and what's next.
The base family, reproduced
45M → 500M trained twice on independent hardware, within 1%. The 1.5B is next — the curve predicts a final loss of 2.37.
The school phase
A teacher-graded curriculum that rewards an honest "I don't know" over a confident wrong answer — models that know when to look things up.
Reasoning depth & world models
Recurrent test-time compute and a world-model interface — the architecture the codebase is built around, activated at the scale where it pays off.
Watch the next model train
One update when a new model finishes — with samples and numbers. Nothing else.