RTX 5090 laptop owner asks: best abliterated Qwen 3.6 vs 3.8 27B quants for 24GB VRAM
ThomasAger · reddit · 2026-09-02
A Reddit user on an RTX 5090 Laptop (24GB VRAM, no offloading) is looking for the fastest abliterated (safety-removed) quantized builds of Qwen 3.6/3.8 27B for instruction-following with thinking disabled, noting community claims that 3.6 performs better without thinking. Some MTP variants run no faster than non-MTP builds for him; he currently uses Huihui-Qwen3.8-27B-abliterated-NVFP4-GGUF and finds abliterated models smarter for his use cases.
More from Models
- Moonshot Deprecates moonshot-v1, Recommends Kimi K3 — emmanuelvivier · 2026-09-02
- Alibaba Unveils Qwen3.8-Flash-Next Previewing Qwen4 Architecture — emmanuelvivier · 2026-09-02
- Ling-3.0-Flash Deployment: Why Total and Active Params Both Matter — Sitkin_Marrel · 2026-09-02
- How good is Qwen 3.8 Flash Next for creative writing? — No_Algae1753 · 2026-09-02
- Fable 5.1: Claimed as Best Model for Coding and Agents — udmrzn · 2026-09-02
- Debate: Reasoning Models Defined by Inference Graph Depth, Not Layer Count — gleech · 2026-09-02