Home › Evidence › Records › cfg-0133
⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.

cfg-0133

backfilled
citable URL: https://halobench.com/records/cfg-0133/ — this address never moves; the anchor /records/#cfg-0133 keeps resolving
modelQwen3.6-35B-A3B-MTP-UD-Q4_K_M.gguf · UD-Q4_K_M · rev 0b21525e972670ed59e1812e170b27c26355381f0656ecc4e25617ece7dac58b
engineggml-org/llama.cpp 7077abbe14c510cb829c93a1328c2815b5805ebd · ROCm0/gfx1151 · host aihydra (igpu)
flagsHO-009 MTP n_max=1 Stage A fingerprint: same as cfg-0132 plus native model MTP/NextN path with --spec-draft-n-max 1. No ngram/external drafter, no KV quant, no proxy/llama-swap, no n_max>=4.
templatenot recorded at test time
treeupstream — stock
aged evidence — reconstructed from the archive. Stage A activation/safety arm only. Positive draft counters were admitted, but n1 was not promoted as the best clean arm.

Capability basis

measured on this config

Runs on this config (2)

rundatekindsuitemetrics
run-05232026-08-20guardmtp-context-proofdecode_tps=68.096503
run-05252026-08-20performancestageA-c32768decode_tps=68.09650276001041