Alibaba Releases Open-Weight Qwen 3.8 Models

Alibaba's AI division, Qwen, has officially released the open weights for its new Qwen 3.8 model family under the permissive Apache 2.0 license. The release features two primary models: the core Qwen3.8-27B, a multimodal dense model containing 27 billion parameters, and the significantly larger Qwen3.8-2.4T-A95B, which is designed to operate at the Max performance tier. Both models are now accessible to the developer community on Hugging Face and ModelScope.

Despite its relatively compact size, the Qwen3.8-27B model reportedly outperforms the older, larger Qwen3.7-Plus in both programming and general office productivity tasks. The development team highlighted the model's enhanced agentic capabilities, noting that it can plan tasks more independently and execute them with greater reliability. Additionally, the model features a flexible thinking mode that is enabled by default but can be toggled on a per-query basis to optimize processing.

For practitioners handling large datasets, Qwen 3.8 offers substantial context handling. It natively supports up to 262,000 tokens of context, which can be scaled up to one million tokens using the YaRN interpolation method. This massive window allows the model to ingest and analyze extensive inputs, including diagrams, documents, and even multi-hour video files alongside standard text and images.

The open-source Apache 2.0 licensing means developers can freely integrate, modify, and commercialize these models without restrictive proprietary barriers. For those who prefer managed infrastructure, Alibaba plans to launch a hosted version of the models featuring the full one-million-token context window on its Qwen Cloud platform in the near future. This combination of open-weight flexibility and cloud scalability provides a highly competitive alternative for enterprise AI integration.