Same model, 30 points apart
Eleven teams, one benchmark, the same model families. The best and worst systems running identical models finished 30.4 points apart, and everything that opened that gap sat outside the weights.
Aug 18, 2026 · 12 min read
Read article →