Alibaba's Qwen team has released the open weights for its new Qwen3.8 generation, headlined by a dense 27-billion-parameter model that the team claims outperforms the larger Qwen3.7-Plus in coding and office tasks.
The core model, Qwen3.8-27B, is a multimodal vision-language model that processes text, images and video natively — including diagrams, documents and multi-hour footage. It handles up to 262,000 tokens of context out of the box and can scale to one million tokens using the YaRN extension method. A flexible thinking mode is enabled by default but can be toggled per query.
According to the team, the biggest gains come in agentic behavior: Qwen3.8-27B plans more independently, completes multi-step tasks more reliably, and is built for developers who want frontier-class automation on consumer-grade hardware. The weights ship under the permissive Apache 2.0 license and are available on Hugging Face and ModelScope.
On the same day, Qwen also published the weights for Qwen3.8-2.4T-A95B, the massive MoE model behind the Max tier, which was described at launch as the largest open-weight release in the company's history. A hosted version of the new models with one million tokens of context is expected soon through Qwen Cloud.
The release lands amid an escalating open-weight race. Chinese labs including Z.ai, DeepSeek and Moonshot have shipped increasingly capable open models this year, and Meta has responded with its own open-weight 'Muse' family. Analysts see the 27B class as the new battleground: models small enough to run locally, yet strong enough to automate real software projects.
