Five of the most capable model families, compared on the architectural choices that are publicly known or widely reported. Sort any column, then flip the view. The skeleton is nearly identical. The parts that still differ are not the architecture.
Families compared
5
Universal traits
3
Converging traits
3
Real differentiators
2
Click a column header to sort.
The drift over time, filterable by family
Move to mixture of experts Move to FP8 training Longer context and efficient attention
Read this as illustrative and public sourced. Every entry is based on public material and informed inference, not confidential detail. Cells marked reported or inferred are cases where a lab has not fully documented the choice, so the value reflects widely reported architecture rather than a disclosed spec. Nothing here states precise unpublished numbers such as parameter counts or training cost. The point is the family level pattern, not any single figure.
When the skeletons match, the moat is the data recipe and the post training, which is exactly what nobody publishes.