Alibaba's Qwen team has opened the model weights for Qwen3.8, expanding access to its latest AI model lineup under the Apache 2.0 license. The release centers on Qwen3.8-27B, a 27 billion parameter multimodal dense model built for text, images, video, coding, office tasks, and agent-style work.
The announcement also includes weights for Qwen3.8-2.4T-A95B, a much larger model designed to operate at the Max level. Together, the models give developers and organizations more ways to evaluate Qwen3.8 outside a purely hosted environment, while a hosted one million token version is planned for Qwen Cloud, Alibaba's AI service.
What Alibaba's Qwen team released
The core model in the release is Qwen3.8-27B. According to Qwen, it is a multimodal dense model with 27 billion parameters, and it outperforms the larger Qwen3.7-Plus in coding and office tasks.
That comparison is notable because the new model is presented as more capable in those practical work categories despite being measured against a larger predecessor. The source does not describe every benchmark or test behind the claim, but it does identify coding and office tasks as areas where Qwen says the model has moved ahead.
The model weights are being released under the Apache 2.0 license. For an open-weight model, the license is a central part of the release because it defines the terms under which people can use the model weights.
Qwen has also released weights for Qwen3.8-2.4T-A95B. The source describes it as much larger than Qwen3.8-27B and built to operate at the Max level.
Long context is a central feature
Qwen3.8-27B natively handles up to 262,000 tokens of context. It can scale to one million tokens using the YaRN method.
In practical terms, long context is about how much material a model can keep in view while responding. A larger context window can matter when the task involves long documents, extended conversations, large codebases, detailed reports, or combinations of material that need to be considered together.
The source also says a hosted version with one million tokens of context will soon be available through Qwen Cloud, Alibaba's AI service. That means the open-weight release and the hosted service are being positioned as related paths into the same broader Qwen3.8 generation.
Multimodal input goes beyond text
Qwen3.8-27B is described as a multimodal model. Beyond text, it can process images and videos, including diagrams, documents, and multi-hour video.
That matters because many real work tasks are not purely text-based. A document may include charts, tables, scans, or diagrams. A video may contain important visual information spread across a long recording. The model's stated ability to process these formats puts Qwen3.8-27B in a category aimed at more than chat-style prompting.
The source also notes a flexible thinking mode. It is on by default, but it can be toggled per query. That gives users a way to adjust how the model approaches individual prompts, without making the setting permanent for every interaction.
Agent capabilities are part of the pitch
Qwen's team is also emphasizing improved agent capabilities. According to the source, the model plans more independently and completes tasks more reliably.
Agent behavior is important when a model is expected to handle multi-step work instead of producing only a single answer. Planning, tracking progress, and carrying out a task reliably are all relevant when the model is used for coding work, office workflows, or tasks that require repeated decisions.
The source does not provide a detailed list of agent benchmarks, but it clearly identifies agent performance as one of the release's intended strengths. Alongside coding, office tasks, multimodal input, and long context, that places Qwen3.8-27B in a role that is meant to support more complex workflows.
Where the Qwen3.8 models are available
Both Qwen3.8-27B and Qwen3.8-2.4T-A95B are available on Hugging Face and ModelScope. Those are the distribution points named in the source for the released weights.
The hosted version with one million tokens of context is planned for Qwen Cloud. The source says it will be available soon, but it does not give a specific date.
The larger picture is a Qwen3.8 release that combines open weights, Apache 2.0 licensing, multimodal input, long-context operation, and a larger Max-level model. For teams comparing AI model options, the key facts are straightforward: Qwen3.8 is now available as open weights, Qwen3.8-27B is the central 27 billion parameter model, and Alibaba is also preparing hosted access through Qwen Cloud.