Alibaba's Qwen Team Releases Qwen3.8 Models with Open Weights under Apache-2.0

Alibaba's AI team Qwen has published the open weights for its Qwen3.8 models under the free Apache-2.0 license. The flagship model, Qwen3.8-27B, is a multimodal dense model with 27 billion parameters that reportedly surpasses the larger Qwen3.7-Plus in many domains, notably programming and office tasks. Qwen also highlights improved agentic capabilities, with the model planning more independently and completing tasks more reliably. It natively processes up to 262,000 tokens of context and can be extended to one million tokens using the YaRN method. Besides text, the model understands images and videos, such as diagrams, documents, or hours-long videos. A flexible thinking mode is enabled by default but can be disabled per request or adjusted in depth. Additionally, Qwen has released the weights for the much larger Qwen3.8-2.4T-A95B model, which is intended for top-tier performance. Both models are available on Hugging Face and ModelScope. A hosted version with one million tokens of context will soon be offered through Qwen Cloud, Alibaba's AI service.