The Qwen3.8-Max-Preview model achieved a SOL Score of 0.761805 with a 12.97x average speedup, leading all submissions in optimizing production-level LLM inference kernels toward NVIDIA B200 hardware limits. Alibaba claims the upcoming open-weight Qwen3.8, with 2.4T parameters, could be one of the strongest models available, second only to Fable 5.