clm-0097
measured-heremed ●●○
citable URL: https://halobench.com/records/clm-0097/ — this address never moves; the anchor /records/#clm-0097 keeps resolving
HO-009 measured reviewer-admitted MTP activation/safety/performance evidence for the genuine Qwen3.6-35B-A3B-MTP UD-Q4_K_M artifact on aihydra. Native MTP draft counters activated for n_max 1/2/3 in Stage A; n_max=2 was the best clean admitted arm and improved server /completion decode throughput versus the plain f16/f16 baseline through d204800 under the exact recorded runtime/config boundary.
verified 2026-08-20 · volatility medium
Note — the record's own working
Scope boundary: MTP activation/safety/performance only for /home/aihydra/bench-results/ho009-qwen36-35b-a3b-mtp-refinement-r2. This is not a full capability result, production recommendation, quality equivalence claim, KV-quant conclusion, or inheritance from invalid 27B-MTP evidence. No n_max>=4 safety claim is admitted. n2 d262144 server-perf is preserved as a dropped HTTP 400 context-fit/refusal cell with no substitution. Energy is joined to Home Assistant wall-meter records as Wh total only for every reviewed window.
Cited by — computed at build time, never stored
candidate gate history qwen36-35b
docs benchmark-protocol