Qianwen First Opened Max Level Flagship Weight, Qwen3.8 Officially Available for Download

According to DataBird Beating Monitor, Alibaba has officially released the Max-level model weights of Qwen 3.8, which have been deployed on Hugging Face and ModelScope. The model is named Qwen3.8-2.4T-A95B, with a total of 2.4T parameters and 95B activations per inference. This marks the first time Qwen has released Max-level flagship weights.

However, the open-source version differs slightly from the cloud-based Qwen3.8-Max. The official model card specifies that the open-source version natively supports 262K contexts, expandable to around 1M, and allows for inference intensity adjustment. The current version only supports textual input and must have Thinking mode enabled. The cloud-based Qwen3.8-Max additionally supports visual input, non-thinking mode, default 1M contexts, and official built-in tools.

The license is no longer Apache 2.0 but a Qwen custom license. Regular users can still download, modify, and deploy the model, but some large-scale commercial uses require separate authorization. Reuters had previously reported that Alibaba plans to introduce additional fees or licensing terms for large commercial users of Qwen3.8-Max.

The previously teased Qwen3.8-27B has not been released simultaneously.