The model fits on a single DGX Spark and runs at 30 - 40 tokens per second, beating open-weight systems near 10x its size.

We over-provision because we assume scale is the only lever, but they pulled persistence and backtracking out of a small model instead.
If a team under 70 people can build infra this tight, what does that say about how lean the rest of us builders and operators could run if reproducibility were the first feature instead of the last?