Qwen 3.8 27B open model matches frontier, runs locally

Alibaba's Qwen 3.8 27B, an Apache 2 licensed 27B vision-capable LLM, scores 52 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Luna (max) and one point behind GLM-5.2 and DeepSeek V4 Pro. It runs locally on an RTX 5090 at up to 200 tk/s, but defaults to 'xhigh' reasoning effort causing overthinking.