Uncensored GLM-5.3-Flash weights released, refusal rates drop to 11%
lipeng0820 · x · 2026-08-29
OrcaRouter released uncensored weights for GLM-5.3-Flash (320B params, 18B active) in native FP8, eliminating the need for LoRA or jailbreak prompts. Evaluations show a significant drop in refusal rates across malicious benchmarks: MaliciousInstruct (96%→11%), JailbreakBench (93%→12%), and others, while benign over-refusal also decreased. Experiments suggest part of the alignment is not mediated by a single linear refusal direction.
More from Models
- MiniMax H3 Max Criticized for Extreme Censorship on Fal — MrUtterNonsense · 2026-08-29
- User Predicts Qwen-4-27B Will Be a Game Changer — Steus_au · 2026-08-29
- AlayaWorld Tops World Model Rankings as First Open-Weights Leader — Obvious_Set5239 · 2026-08-29
- Terminal-Bench 4.0: GLM-5.3 Surpasses GPT-5.6 in New Ranking — eyishazyer · 2026-08-29
- Minimax H3 limits: Identity leaking and quality degradation — Kooky-Mode3047 · 2026-08-29
- Qwen3.8 27B runs at 50 tok/s with 100k context on 16GB GPU — qaf23 · 2026-08-29