2月9日，在全球最大AI开源社区HuggingFace的开源项目页面中，最新出现了Qwen3.5并入Transformers 的新PR（提交代码合并申请）。业内猜测，阿里千问新一代基座模型Qwen3.5发布在即。该模型采用了 全新的混合注意力机制，并且极有可能是原生可实现视觉理解的VLM类模型，或将开源至少2B的密集 模型、35B-A3B的MoE模型。 点评：Qwen 3.5模型代码现身开源社区，预示着其将延续并强化开源战略。选择开源至少2B的密集模 型与MoE模型，或在于通过技术开放快速获取开发者生态、构建应用层事实标准。在基础模型能力逐渐 接近的竞争格局下，生态的广度与活跃度已成为比单纯参数规模更核心的壁垒。 NO.3 擎天租2月首周订单环比增长约30% ｜2026年2月10日星期二｜ NO.1 贾国龙卸任西贝小牛法定代表人 2月9日消息，北京西贝小牛餐饮管理有限公司2月5日完成工商变更，公司名称已调整为北京贝牛餐饮管 理有限公司。同时，西贝集团创始人贾国龙卸任法定代表人、经理及董事职务，转由何海宾接任。 点评：结合贾国龙近期"回归一线、聚焦主业"的公开表态，此举可视为其作为实控人进行的一次清晰的 战略收 ...

Core Insights - Alibaba is leveraging the Qwen large model foundation to comprehensively reshape its business and accelerate the construction of B-end ecological barriers through an open-source strategy and strong performance [2][42] - The company is committed to increasing capital expenditure to meet strong demand for computing power, with cloud revenue continuing to grow significantly, validating the closed-loop logic of "infrastructure investment - technology iteration - commercial monetization" [2][42] AI Model Development - Alibaba has been early in the AI model layout, with its flagship Qwen series iterating three major versions and multiple minor versions in just over two years, covering vertical scenarios such as text, mathematics, code, and multi-modal applications [3][43] - As the only major player adhering to an open-source strategy, Alibaba has accelerated model iterations since 2024, narrowing the capability gap with overseas models and is expected to surpass closed-source models in the B-end market [3][43] Model Capabilities and Updates - Alibaba's AI model layout has achieved "full size," "full modality," and "multi-scenario" coverage, with the first trillion-parameter multi-modal model M6 released in March 2021, and subsequent expansions in parameter size [4][43] - As of October 2025, the Qwen model series has iterated three major versions and multiple minor versions, covering model sizes from 0.5 billion to one trillion parameters [4][43] - The Qwen series has open-sourced a total of 357 models, with a significant acceleration in update frequency since 2024, including 71 models in the first half of 2024 and 120 in the second half [5][45] Competitive Positioning - Alibaba's model capabilities rank in the global first tier, with the gap to leading overseas firms reduced from over six months to approximately three months [12][51] - The Qwen3 235B model, released on July 22, 2025, is comparable to DeepSeek-V3.1 Terminus, while the Qwen2.5 Instruct-72B model was the first domestic model to exceed an overseas model in the open-source arena [12][51] Market Strategy and Ecosystem - The company has formed a complete layout for both B-end and C-end markets, with open-source serving as the foundation for business and product support [35][74] - As of the 2025 Yunqi Conference, the Tongyi series models have achieved 600 million downloads and served over 1 million customers, with the Mota community further enhancing the open-source ecosystem [35][74] Future Outlook - The focus will be on the official release of Qwen3-Next (equivalent to Qwen3.5) and optimizing vertical models based on it, with Qwen4 expected to be released in Q2 2026 [21][60] - The ongoing open-source strategy is anticipated to maintain the frequency of model updates, further enhancing capabilities while narrowing the time gap with leading overseas models [21][60]