Stage-3 OPD @ 600M generated tokens — the H=16K MiMo branch extended from the 400M control's pre-decay state (340.02M tokens, step 2265) to 600.03M tokens (WSD schedule: flat 1e-5 to ~536.8M, 482-step linear decay to 0). Snapshot final (step 3974) of run mimo-wsd-rugfix-h200x3-ext400-decay0-v1 (W&B project mimo-gdn-linearization). Evals in full_eval/: RULER-lite n=100 seed 0: niah_multikey@32768 0.84, niah_single 1.00 at all lengths, vt 0.966@32k; AIME'24 avg@8 (think@32k) 0.5125, AIME'25 0.3583; MATH-500 think 0.928; GSM8K@1024 strict 0.52995 (the dip vs 400M is a suspected 1024-token generation-cap artifact — chains lengthened through the decay tail; AIME/MATH/ RULER all improved — pending a @4096 confirmation rerun).

Part of the MiMo second-model arm of a study converting full-attention models into GDN (gated DeltaNet) hybrids (uniform 1:4 retention) and recovering capability via staged distillation + on-policy distillation. Teacher/base: MiMo-7B-RL-0530.

NOTE: these checkpoints use the custom mimo_gdn architecture — register it before loading (import src.models.mimo_gdn from the project repo, or use the load_causal_lm helper in src/eval/sample_generations.py). Bare AutoModelForCausalLM will not work. vLLM serving needs src.models.vllm_mimo_gdn.register().

Sibling checkpoints: mimo-7b-gdn-hybrid-200M-OPD (base control), -400M-OPD (H=16K control final), and -800M-OPD from the H=32K extension when it completes.