| # Optimization results | |
| | Variant | Checkpoint | Schedule | Tier 6 accuracy | Tier 6 time (MPS) | | |
| |---|---:|---|---:|---:| | |
| | Published v8 | 1,887,610 B | reduce a + reduce b + multiply | 100/100 | 27.02 s | | |
| | Compact weights | 948,196 B | reduce a + reduce b + multiply | 100/100, identical outputs | 34.59 s* | | |
| | Compact + short schedule | 948,196 B | reduce a + multiply by streaming b | 100/100, identical outputs | 20.45-21.31 s | | |
| | Compact + short schedule + static-channel reuse | 948,196 B | reduce a + multiply by streaming b | 100/100 | 19.07 s | | |
| `*` This single MPS timing is noisy and does not imply bfloat16 storage makes | |
| inference slower. The compact tensors are loaded back into fp32 parameters; | |
| storage precision affects artifact size, not the execution dtype. | |
| The short-schedule variant also scored 100/100 on each public Tier 1 through 7 | |
| (700/700 total) and passed the official static checker. Full Tier 8-10 and | |
| held-out adversarial validation remain release blockers. | |