STRUCTURE IS THE DESIGN
Models that can stand the test of space and time.
The Geometric Rotation Transformer stores weights as rotations and tokens as phases — roughly half the numbers, stable before the first training step. Code and benchmarks land as the training run completes.
HOW WE WORK
We pick structure over scale.
We design model architectures from first principles — geometric constraints instead of raw parameter count, so every number in the model has a job.
Our first architecture, the Geometric Rotation Transformer, replaces dense weight matrices with rotations. Same attention, same depth — a different primitive underneath. We train it, measure it, and publish everything: code, configs, weights.
EARLY ACCESS
Be first to get GRT.
Get the code, weights, and benchmark report the day GRT training completes.