On August 13, the Alibaba Cloud ModelScope community made an announcement: the Alibaba Qwen team has officially rolled out the model weights for Qwen3.8-2.4T-A95B. This move signifies the inaugural open-source release of a model at the Qwen-Max level.

The Qwen3.8-2.4T-A95B model is truly remarkable, boasting a staggering 2.4 trillion parameters. It employs a Mixture of Experts (MoE) sparse architecture. In this architecture, only 95 billion parameters are activated for each Token. Moreover, it inherently supports a context window of 262,144 Tokens, and this capacity can be expanded up to 1,010,000 Tokens.

The official version, Qwen3.8-Max, which is based on Qwen3.8-2.4T-A95B, was launched on August 3. It comes with visual input capabilities and offers default support for processing contexts that are up to 1 million in length.