Alibaba’s open-weight Qwen3.8-Max takes on long-horizon AI tasks with 2.4 trillion parameters

Alibaba has unveiled Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters per query, designed for autonomous, multi-day work rather than one-off responses. The company plans to release its weights on Hugging Face and ModelScope next week, making it the first Qwen-Max model to become publicly available. Alibaba presented case studies in coding, research reproduction, chip design, and simulated e-commerce to demonstrate the model’s long-horizon capabilities. In the company’s reports, Qwen3.8-Max independently built a command-line tool over 16 days, reproduced and improved results from a research paper over about five days, reduced a cryptographic circuit from 8,298 to 678 logic gates, and more than quadrupled starting capital in a simulated year of online retail. These results are company-reported and have not yet been independently verified. The model also supports long documents and videos, and Alibaba introduced RecreationBench for testing whether models can reconstruct applications through interaction alone. In Alibaba’s internal benchmark comparisons, Qwen3.8-Max scored highly across coding, agentic, reasoning, and multimodal tasks,|