Qwen3.8-Max Posts Benchmarks and Promises Weights Next Week

Alibaba’s 2.4-trillion-parameter mixture-of-experts model activates about 95B per request, handles text, images and video, and supports a 1M-token context. Published scores include 93.0 on PaperBench, 82.8 on IFBench, 82.3 on MMMU-Pro, 86.1 on OSWorld-Verified and 90.4 on VideoMME with subtitles. The third-party number worth more than any of those: second place in Vision Arena, fifth in Text Arena. It is live on QwenCloud now, with weights announced for Hugging Face and ModelScope next week. If that release lands, it resets the open-weight ceiling. The licence terms have not been stated.