Home › Evidence › Records › gpt-oss-120b

gpt-oss-120b

rejected
citable URL: https://halobench.com/records/gpt-oss-120b/ — this address never moves; the anchor /records/#gpt-oss-120b keeps resolving
gpt-oss-120b @ Q4_K_M · kind model · engine igpu · tier candidate
runs on aibeast · config cfg-0003
⌁ current state rejected · ⌁ days in production 0

Lifecycle — append-only

2026-06-27
candidate · run-0003
2026-06-28
rejected — dec-0001 · run-0003
The migration's planned upgrade and Stage-1 pick, reversed by harder testing. Fastest reactive model (~5 s) but placed LAST on the blind conceptual set: 3/8, mean 1.38. Its terseness read as under-delivery rather than restraint — blind judges independently described answers as "non-answer", "fragment of internal musing", "a JSON list of fictional tool calls". See clm-0004 for the fingerprint and the fairness conditions this result depends on.