Alibaba's Qwen3.8-2.4T-A95B, a massive 2.4-trillion-parameter MoE model, launched with day-zero adap...
The AMW Read
Alibaba's open-weight release with day-one multi-chip adaptation (including domestic Chinese silicon) is a major milestone for open-model portability and compute diversity, though it builds on FlagOS's ongoing work.
Alibaba's Qwen3.8-2.4T-A95B, a massive 2.4-trillion-parameter MoE model, launched with day-zero adaptation across nine AI chip platforms, enabled by BAAI's FlagOS open-source stack. The model, Alibaba's largest open-source release yet, brings Qwen-Max-level capabilities to the open ecosystem, with 95 billion active parameters, a 262,144-token native context (expandable to 1,010,000), and support for BF16/FP8 weights. FlagOS teams completed multi-chip adaptation, precision alignment, and deployment validation on chips from T-Head, NVIDIA, Moore Threads, Huawei Ascend, MetaX, KunlunXin, Hygon, Qingwei Intelligence, and Suiyuan, offering BF16, FP8, and INT8 precision variants for each platform. This effort builds on FlagOS's track record since February, now covering 12 open models from seven major teams across ten chip types, compressing new-model-to-multi-chip-ready time to under 24 hours.
