qwen35-122b-a10b

Displaced the benchmark winner on a deliberate override of the data: the point of the exercise was to find what Warden could become, and the 122B bets on stronger cross-domain reasoning showing up in real proactive work. Later reinforced by the MTP variant (+49% decode) settling the choice. Reference-only in the benchmark — it matched the 35B on reactive speed despite 10B active vs 3B, which is what exposed reasoning-token length rather than decode speed as the latency driver. Ran in production on the original aibeast until that box's board-level power failure (inc-0005); it remained the documented switch-back target through the outage and was formally displaced on 2026-09-02, when the rebuilt aibeast came up serving Qwen3.8-Flash-Next (agention ROCmFP4-FAST-v2) as the production model (clm-0128). Kept visible as the switch-back reference, not current production.