On August 13, Alibaba made its ultra-large-scale Mixture of Experts (MoE) model, Qwen3.8-2.4T-A95B, available as open-source and launched it on the supercomputing internet. This model features an impressive total of 2.4 trillion parameters and employs a sparsely activated hybrid expert architecture, ensuring performance that rivals the top models in the industry. By utilizing the 10,000-card supercluster at the core nodes of the supercomputing internet, enterprises and developers can directly access model images that have been adapted using the unified software stack of Zhongzhi FlagOS and domestic heterogeneous accelerator cards within the AI community. This facilitates the swift deployment of AI application services. Additionally, Qwen3.8-2.4T-A95B has been successfully adapted to nine AI chips, eight of which are domestically produced, demonstrating the scalable implementation capability of achieving 'one-time development, rapid multi-chip adaptation' for advanced models based on a unified and open system software stack.
