OrcaRouter 🐳(@OrcaRouter):Someone independently benchmarked 8 different Qwen3.8-27B uncensored/abliterated models. We won. 🐳 Abliterlitics spent ~167 GPU-hours testing the variants across weight forensics, KL divergence, 13 capability benchmarks + 400 HarmBench behaviors. OrcaRouter Qwen3.8-27B-Uncensored: → #1 Judge ASR: 82.2% → Base model: 4.5% → MMLU-Pro: −0.0pp → IFEval: +0.4pp → 4/4 of our model-card claims independently verified Their verdict: “The winner.” The interesting part isn't just removing refusals. It's removing them without destroying the intelligence underneath. Independent analysis: https://t.co/6EM9D5lOmz Weights:

2026.09.07 05:11

Someone independently benchmarked 8 different Qwen3.8-27B uncensored/abliterated models. We won. 🐳 Abliterlitics spent ~167 GPU-hours testing the variants across weight forensics, KL divergence, 13 capability benchmarks + 400 HarmBench behaviors. OrcaRouter Qwen3.8-27B-Uncensored: → #1# Judge ASR: 82.2% → Base model: 4.5% → MMLU-Pro: −0.0pp → IFEval: +0.4pp → 4/4 of our model-card claims independently verified Their verdict: “The winner.” The interesting part isn't just removing refusals. It's removing them without destroying the intelligence underneath. Independent analysis: Weights:

480

轉發到社區

熱門用戶