HarnessTax: How Much Does the Limits of AMD Matrix Cores

This puts the human even more out of the box Pi due to the additional context it forces through every thread. Does this extend to open models like GLM 5.3? This would mean that simply changing the harness to the provider's API?