Two Local GPU Breakthroughs Allow Running Alibaba's Qwen 3.8 Models on Home Hardware

Recent developments for running Alibaba's Qwen 3.8 models on normal home hardware highlight two major local GPU breakthroughs. Strata 0.2.10 and llamAmpere v0.6.3 enabled the 125-billion parameter Qwen3.8-Flash-Next to run at over 141 tokens per second on a single 12GB RTX 5070 with a 94.7% GPU hit rate. Additionally, Bonsai 2 and Ingot-XXVII compressed Qwen3.8-27B by 89.4%, reducing its size to 5.9 GB while retaining 98.2% accuracy. Open article →