Alibaba releases Qwen3.8-2.4T-A95B open weights; NVIDIA serves it on GB300 NVL72

Qwen3.8-Max is a sparse MoE with 2.4T total parameters, 95B active per token, and a 1M-token context. NVIDIA reports Day-0 throughput above 4K tokens/s per GPU on GB300 NVL72 in FP8, with Qwen calling it its first multimodal model above 1T parameters.