Full state (BF16 model shards + FP32 optimizer shards + trainer_state) of the MiMo-7B GDN-hybrid OPD run mimo-wsd-rugfix-h200x3-ext600-decay0-v1 (W&B id mimo-rugfix-h200x3-ext600-decay0-v1-20260813) at step 5204 / 800,014,301 consumed tokens, H=32K. Saved on the Empire 3xH200 7-day reservation (job 16007, WSD OPD scaling ladder). CORRECTED 2026-08-17: earlier upload shipped stale root model shards (branch-init state, 536,959,878 tokens); model shards, index, and snapshot_meta now come from snapshots/final (true 800M final, step 5204, 800,014,301 tokens). Optimizer/trainer state were already correct.

Custom mimo_gdn architecture - register before loading (see project repo).

Source Git commit: e3cc4c0ff85a10b0660af706e80f21bfb252677b