Reddit user ships 'surgical abliterated' 27B red-team model with zero refusals
Least_Dog_8556 · reddit · 2026-10-07
A Reddit user released Qwen3.8-cyber-RedTeam-27B, an 'unconstrained' fine-tune claiming zero refusals via residual-stream orthogonalization of refusal directions, 256K context, hybrid SSM architecture, native FP8 on a single 32GB GPU, and agentic tool-calling — aimed at authorized red-teaming. Note: the naming and architecture claims are unverified (no official Qwen 3.8 exists), and fully uncensored security models raise obvious misuse concerns.
More from Models
- OpenAI reasoning models went from basic arithmetic to decades-old math breakthroughs in two years — daniel_mac8 · 2026-10-07
- Gary Marcus: OpenAI's Math Proofs Rely on Symbolic AI, Vindicating His Stance but Not AGI — GaryMarcus · 2026-10-07
- Ex-Googler suspects Project Astra voice got quantized and downgraded — joannejang · 2026-10-07
- Running quantized Swift 1.5 in 16GB VRAM at 50 tps: can lower quants help agentic coding? — royalflash417 · 2026-10-07
- OpenAI Reportedly Makes Serious Breakdown Toward Riemann Hypothesis — ipeirotis · 2026-10-07
- Jev-as-a-Judge: New Model Boosts LLM Judge Reliability for Agent Evals — omarsar0 · 2026-10-07