EVENT DOSSIER
On August 12, 2026, Alibaba announced through NVIDIA Developer Blog the public weights of its largest open-source model, Qwen3.8-2.4T-A95B (Qwen3.8-Max). This model has a total of 2.4T parameters, with 95B activated parameters per token. It uses a fine-grained MoE architecture, combining full attention and linear attention mechanisms, and supports context windows of up to one million tokens along with corresponding output lengths. Its purpose is to provide the open-source ecosystem with capabilities close to cutting-edge levels.
Alibaba Qwen3.8-2.4T-A95B 发布开源大模型，参数达 24 万亿
阿里巴巴发布其最大开源模型 Qwen3.8-2.4T-A95B（Qwen3.8-Max）的公开权重，该模型拥有 2.4T 总参数和每 token 95B 激活参数。它采用细粒度混合专家（MoE）架构，结合全注意力与线性注意力机制，支持高达一百万 tokens 的上下文窗口及相应输出长度，旨在为开源生态带来接近前沿的能力。