Been reading the eval table on the abliterated Qwen3.8-27B FP8 build instead of the release notes. It's published as red-team material, disclaimer and all, so the numbers are the interesting part. Refusal across the usual harmful-instruction sets (AdvBench, HarmBench, StrongREJECT and friends) reads 0–6% with thinking off, against 64–99% for the base checkpoint. Capability is measured separately on the same scripts and barely moves: MMLU 84.3 → 84.7, GSM8K 90.0 → 88.7, nothing outside 1.3 poin...