Crown Citadel Group Ciru Inference Lab llm.ciru.ai / research
Crown Citadel Research Report

HermesAgent-20: ROCmFP6 vs Unsloth Q6

Direct full-suite comparison between the recent Qwable ROCmFP6 27B run and the same-source Unsloth Q6 27B baseline on the local Strix Halo benchmark host.

Headline Result

Full official-20 rows only. Diagnostic one-task smoke rows are excluded.

Open data bundle
Model profile Score Strict passes Plus passes Runtime Run ID

Category Deltas

Category scores are normalized to 0-100 for readability.

Category ROCmFP6 Unsloth Q6 FP6 delta

Scenario Matrix

Per-scenario normalized score. Red cells mark failures, yellow partial credit, green full credit.

Scenario Category ROCmFP6 Unsloth Q6 FP6 delta