Old Zhang's AI Learning

Sep 13, 2026 · Artificial Intelligence

Qwen3.8-27B Turbo: Cutting Reasoning Tokens 10x While Hitting 735 ARC-C

A modified Qwen3.8-27B model achieves 735 ARC-C and 882 ARC-E scores while slashing reasoning tokens up to 10x through TURBO thinking reduction, Cold Fusion training with GAIN dynamic scheduling, Heretic+ARA uncensoring, and Dual iMatrix+MTP quantization for consumer GPU deployment.