clm-0120
measured-heremed ●●○
citable URL: https://halobench.com/records/clm-0120/ — this address never moves; the anchor /records/#clm-0120 keeps resolving
A Qwen3.8-27B ROCmFPX-Q4 plus DFlash2 server screening sample reported 52.2 t/s on a varied code prompt and 31.5 t/s on a varied prose prompt. A separate five-task airline smoke completed 5/5 with mean reward 1.000. These are bounded source observations, not a production-throughput, quant-equivalence, or full capability verdict.
Note — the record's own working
The source is an operator-directed autonomous handoff, not an hbrunner run. The served-path samples use distinct prompts, but the handoff supplies neither the mandatory plain/llama-bench floor pair nor a house guard, artifact SHA256s, an energy window, or a blind-judge artifact for prose. The varied-prompt caveat is load-bearing: repetitive prompts inflate speculative acceptance. Do not convert this screening sample into a production figure.
Cited by — computed at build time, never stored
model pages qwen38-27b
candidate gate history qwen38-27b