Abliterated Qwen3.8-27B Loops on Unrelated Tool Calls in VS Code Copilot
fbms2 · reddit · 2026-08-19
A user reports that huihui's qwen3-8-27b abliterated build (Q6KL) misbehaves in VS Code Copilot at both xhigh and low reasoning: it keeps making unrelated tool calls and outputs nonsense in its thinking. The stock Qwen3.8-27B-Q6K shows no such issue, suggesting the abliteration process broke its reasoning.
More from Models
- Grok Bot Ran a 12-Hour Data Analysis Nonstop, Handling API Rate Limits Itself — Daniel_Farinax · 2026-10-16
- One year of Recursive Language Models: GPT-5-mini beats GPT-5 by 110% on 132k-token tasks, cheaper too — a1zhang · 2026-10-16
- Users Report Claude Models Suddenly More Compliant After Forced Compaction — RealSharpNinja · 2026-10-16
- Prediction: Opus 5.5-level models on laptops by 2027, backed by Qwen 27B data — cephaloform · 2026-10-16
- Users Say Krea 2 Output Degraded Sharply Since Krea 3 Beta Launched — szywis · 2026-10-16
- LLM-as-a-Verifier lifts Nemotron 3.5 Super by 9-10 points on agentic benchmarks — simonguozirui · 2026-10-16