AnalysisAI ModelsSeptember 12, 2026

UkisAI post-trains Qwen 3.8 27B to cut thinking 58.3%

UkisAI's Swift-Qwen3.8-27B penalizes tokens linked to overthinking, then restores accuracy via on-policy distillation, cutting thinking tokens 58.3% at 1.95x speed while matching xhigh accuracy. The GGUF build has 92 likes and 16,624 downloads.

3 sources

More stories today

Open the live feed