阿里宣布开源 Qwen3.8-27B 大模型，所有开发者、科研机构和企业可自由下载、部署和使用[1]。该模型是原生多模态稠密模型，支持 262K 原生上下文，通过 YaRN 技术可外推至 1M Tokens 上下文[1]。
与前代 Qwen3.6-27B 相比，Qwen3.8-27B 在编程和办公场景的性能大幅提升，表现超越 Qwen3.7-Plus[1]。该模型新增 reasoning_effort 功能，可根据任务难度控制思考深度[1]。27B（270 亿参数）是全球 AI 社区呼声最高的模型尺寸[1]。
Alibaba has announced the open-source release of its Qwen 3.8-27B large language model, making it freely available for download, deployment, and use by all developers, research institutions, and enterprises.[1] The 27-billion-parameter model represents the most requested model size within the global AI community.[1]
The Qwen 3.8-27B is a natively multimodal dense model that supports 262K tokens of native context length, with the capability to extend to 1 million tokens through YaRN technology.[1] Compared to its predecessor Qwen 3.6-27B, the new model demonstrates significant performance improvements in programming and office application scenarios, even surpassing the capabilities of Qwen 3.7-Plus in these domains.[1] Additionally, the model introduces a reasoning_effort feature that allows users to adjust the depth of reasoning based on task complexity.[1]
