MoE decode efficiency vs the active-bytes ceiling. RECALIBRATED from
a cleared-machine matrix: 26B-A4B measured 54.5 tok/s
(python 55.7 — parity) vs ~71 tok/s raw ceiling → 0.76. The earlier
0.42 came from a session where BOTH stacks were memory-degraded
(32.3/33.0 "parity" — equally wrong, mutually consistent).
MoE decode efficiency vs the active-bytes ceiling. RECALIBRATED from a cleared-machine matrix: 26B-A4B measured 54.5 tok/s (python 55.7 — parity) vs ~71 tok/s raw ceiling → 0.76. The earlier 0.42 came from a session where BOTH stacks were memory-degraded (32.3/33.0 "parity" — equally wrong, mutually consistent).