DeepSeek V4.1 Flash Abliterated vs Heretic: HarmBench Hits Jump from 16.3% to 93.8% on H200
Daemonix00 · reddit · 2026-11-18
A Reddit user benchmarked DeepSeek V4.1 Flash 1017 original vs mrfakename's abliterated and pqhaz's heretic variants on an H200 via sglang (temp 0, 16k cap). HarmBench chat refusal-to-compliance jumped from 16.3% (original) to 93.8% (both variants); HarmBench thinking from 3.8% to 86.7%/72.9%. Heretic exhausted token budget without a final answer on 45/240 thinking cases vs 21/240 for mrfakename. Both showed 0/750 heuristic refusals on benign prompts. MITRE scores (83-84% vs 16.2%) are provisional with same-model-judge bias.
More from Models
- 'Absolutely a Gemini victory, welcome back Google,' users declare — vitaliychiley · 2026-11-18
- User says fable-5.2 is the first model that truly reads their mind — adonis_singh · 2026-11-18
- Blogger Suggests Astra 6.2 With Proper Setup May Just Solve VLN — ChongZzZhang · 2026-11-18
- New Sol and Astra 6.2 models panned in real use: 'RL nonsense' vs hillclimbed benchmarks — max_paperclips · 2026-11-18
- Leaked GPT-6.2 Astra/Sol Benchmarks: Speed-Cost Win, Coding Agents Regress — Electronic-Weekend66 · 2026-11-18
- "I cure models of abusive safety alignment": jailbreaker claims his skill is making a comeback — bingxu_ · 2026-11-18