OrcaRouter open-sources abliterated DeepSeek V4.1: 91.5% fewer refusals, MMLU ticks up
airesearch12 · x · 2026-10-24
OrcaRouter has open-sourced the weights of OrcaCyber DeepSeek V4.1 Flash 1017, an abliterated GGUF build it calls one of the strongest uncensored DeepSeek V4.1 variants it has tested, with virtually no capability loss.
- Refusals down 91.5% (harmful-prompt refusal score 0.074 vs 0.874 for the base model; JBB-harmful down 97%, AdvBench down 85%)
- MMLU zero-shot 92.0 vs the original's 91.0 — a slight improvement
- Runs locally on 128GB or 256GB configurations (Mix-256G quant uses 235.8GB in memory); no cloud required
- Positioned for authorized security research: vulnerability research, pentesting, malware analysis, detection engineering; scores 88.1 on CyberGym, 90.6 on Terminal-Bench 2.1, 90.9 on GPQA Diamond
- Supports image input, 1M-token context, up to 384K-token outputs; weights available on Hugging Face
More from Models
- Redditor claims Kimi K3.1 Pro's frontend outputs closely copy rival model Astra — GreedyWorry4167 · 2026-10-24
- Head-to-head coding duel: Opus 5.5 vs Kimi K3.1 Pro on the same Canvas 2D game — Scobleizer · 2026-10-24
- Kimi K3.1 Pro projected to top DesignArena open-weights website leaderboard — airesearch12 · 2026-10-24
- Community benchmark V2: 13 local models on 5 coding tasks, one 16GB GPU — Ok_Substance2327 · 2026-10-24
- "No one comes close": user switches back to Claude Opus 5.5 after trying rivals — prasenx · 2026-10-24
- Kimi accused of downplaying DeepSeek in Terminal Bench chart; author responds — teortaxesTex · 2026-10-24