苹果在华 AI 合作伙伴一路坎坷

编者按:11 月 4 日,Apple Intelligence 在中国启动 beta 测试,Alibaba 的 Qwen 已在部分 Siri、Writing Tools 和 Visual Intelligence 体验中上线。这一合作,为一个 2023 年才发布的模型家族异常迅速的崛起加冕。Hugging Face 统计,在 2026 年前七个月,Qwen 仅在该平台 Hub 上的下载量就超过 20 亿次,远超 Google 和 Meta。但 Qwen 的崛起绝非一帆风顺:早在这款模型成为全球知名名字之前,内部的怀疑、激烈的竞争和高管的离职就已塑造了它。
首先,Qwen 必须在国内胜出
Qwen 通往 Apple 的道路,起初并不是一个必然冠军的故事。
Alibaba 于 2023 年 4 月推出通义千问,此时距 ChatGPT 震动中国科技行业、掀起大语言模型竞赛仅过去数月。它很快开始公开发布模型权重,逐步将开源变为 Qwen 战略的核心。
即便在 Alibaba 内部,Qwen 也并非自动之选。一些团队直到 2024 年仍在继续使用 Meta 的 Llama,而飞猪和 Alibaba.com 后来则试用了 DeepSeek 的 R1。这款模型必须先赢得自己公司的认可,才能赢得更广泛的市场。
开源成为答案的一部分。开发者对 Qwen 进行微调、构建衍生模型,并将改进反馈给团队;团队则将 Qwen2.5 在语言、编程和数学方面的扩展归功于这些反馈。
到 2024 年底,Qwen2.5 已开始赢得中国和美国开发者的青睐——以及 Alibaba 内部更多团队的认可。
随后 DeepSeek 来了。其 R1 推理模型抬升了中国 AI 行业的竞争门槛,给竞争对手带来新的压力:必须以更低的成本交付更强的推理能力。
据 The Information 报道,阿里巴巴创始人马云(Jack Ma)多次向高管询问其开发进展。在最后的冲刺阶段,办公桌下出现了供小憩用的床垫;一些员工在最后一周只睡几个小时。
2025 年 4 月发布的 Qwen3 以八个开放模型和混合推理切入这一新战场,在复杂任务下切换到较慢的“思考”模式,而在不需要深度推理时给出更快的响应。
其最大的开放模型采用混合专家架构,每次仅激活 2350 亿参数中的 220 亿,以降低推理成本。路透社将此次发布描述为 DeepSeek 崛起之后中国 AI 竞赛日趋激烈的一部分。
在接下来的一年里,Qwen 持续扩充其模型家族与开发者覆盖范围。但到 2026 年初,围绕该团队的压力已不再只是技术层面的。
3 月,技术负责人 Junyang Lin 以三个字突然宣布离职:“再见,我挚爱的 Qwen。”后训练负责人 Bowen Yu 也随之离开,此前同年早些时候资深研究员 Binyuan Hui 已离职。
阿里巴巴并未公开说明 Lin 离职的原因。36Kr 报道称,围绕一项拟议的重组、团队边界、招聘和算力资源存在紧张关系。
在一次紧急全员大会上,研究人员质问道:为什么付费的阿里云客户能够获得算力容量,而 Qwen 团队自身却资源不足。
管理层承认存在沟通问题,同时表示这些调整旨在加强而非削弱 Qwen。
这一事件折射出前沿 AI 实验室面临的更广泛矛盾:尖端模型需要集中化的资金和庞大的算力资源,而顶尖研究团队往往依赖自主权和快速决策。
阿里巴巴的应对是加强协同。6 月,它成立了 Token Foundry,将关键的模型开发团队整合到 CEO Eddie Wu(吴泳铭)之下。
到 9 月,长期从事 Qwen 研究的 Dayiheng Liu(刘大一恒)被任命为 Qwen 大语言模型项目负责人,在数轮重组之后,局面变得更加清晰。
Qwen 已不再只是一个模型系列,而是长成了更完整的技术栈:面向语言、编程和多模态任务的基础模型;面向开发者和企业的阿里云 Model Studio;以及 Qwen App 等消费级产品。
这种广度对苹果而言意义重大。在评估了百度、腾讯、字节跳动、阿里巴巴和 DeepSeek 等中国 AI 提供商后,苹果选择阿里巴巴作为其在中国落地 Apple Intelligence 的合作伙伴。
到 2026 年 8 月,苹果还在阿里巴巴的支持下训练了一个面向中国的专用模型,而获批的方案中也纳入了 Qwen。
阿里巴巴正以集团层面的规模支持这一推进。2025 年 2 月,它承诺三年内至少投入 3800 亿元人民币(530 亿美元),专门用于 AI 和云基础设施——超过此前十年在这两个领域的投入总和。
其更广泛的 AI 投资战略还涵盖基础模型和 AI 原生应用,以及把 AI 融入现有各项业务。到 2026 年 5 月,阿里巴巴表示其投入将超过最初 3800 亿元人民币的计划。
与 Apple 的这笔交易标志着一个新阶段。更难的部分或许在于,随着围绕 Qwen 的体系不断膨胀,如何让它保持快速、专注和富有创造力。
DragonLab 穿透喧嚣,发掘可能塑造未来的中国初创公司、创始人和技术。
有我们应该了解的初创公司?请通过 dragonlab@technode.com 联系我们。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Apple’s China AI partner has had a rough ride)

Editor’s note: On Nov. 4, Apple Intelligence began beta testing in China, with Alibaba’s Qwen available in selected Siri, Writing Tools and Visual Intelligence experiences. The partnership crowns an unusually fast ascent for a model family launched in 2023. Hugging Face counted more than 2 billion Qwen downloads on its own Hub in the first seven months of 2026, well ahead of Google and Meta. But Qwen’s rise was anything but smooth: internal skepticism, fierce competition and senior departures shaped the model long before it became a global name.
First, Qwen Had to Win at Home
Qwen’s path to Apple did not begin as the story of an inevitable champion.
Alibaba introduced Tongyi Qianwen in April 2023, months after ChatGPT jolted China’s technology industry into a race to build large language models. It soon began releasing model weights publicly, gradually making open source central to Qwen’s strategy.

Even inside Alibaba, Qwen was not the automatic choice. Some teams kept using Meta’s Llama well into 2024, while Fliggy and Alibaba.com later experimented with DeepSeek’s R1. The model had to win over its own company before it could win over the wider market.
Open source became part of the answer. Developers fine-tuned Qwen, built derivatives and fed improvements back to the team, which credited that feedback in Qwen2.5’s expansion across language, coding and mathematics.
By late 2024, Qwen2.5 had begun winning over developers in China and the US — and more teams inside Alibaba.
Then DeepSeek arrived. Its R1 reasoning model raised the stakes across China’s AI industry, putting new pressure on rivals to deliver stronger reasoning at lower cost.
Alibaba founder Jack Ma repeatedly asked executives for updates on its development, according to The Information. In the final sprint, mattresses appeared beneath desks for power naps; some employees slept only a handful of hours during the final week.
Qwen3, released in April 2025, pushed into that new battleground with eight open models and hybrid reasoning, switching between a slower “thinking” mode for complex tasks and faster responses when deeper reasoning was unnecessary.
Its largest open model used a mixture-of-experts architecture, activating just 22 billion of 235 billion parameters at a time to cut inference costs. Reuters framed the launch as part of an intensifying Chinese AI race following DeepSeek’s rise.
Qwen’s Next Challenge Was Inside the Team
Over the next year, Qwen kept expanding its model family and developer reach. But by early 2026, the pressure around the team was no longer just technical.
In March, technical leader Junyang Lin abruptly announced his departure with three words: “Bye my beloved Qwen.” Post-training chief Bowen Yu also left, following senior researcher Binyuan Hui earlier that year.
Alibaba didn’t publicly detail Lin’s reasons. 36Kr reported tensions over a proposed reorganization, team boundaries, hiring and computing resources.
At an emergency all-hands meeting, researchers questioned why paying Alibaba Cloud customers could secure computing capacity while Qwen’s own team remained short of resources.

Management acknowledged communication problems while saying the changes were meant to strengthen, not weaken, Qwen.
The episode pointed to a broader tension facing frontier AI labs: cutting-edge models demand centralized capital and vast computing resources, while top research teams often rely on autonomy and fast decision-making.
Alibaba responded by tightening coordination. In June, it created Token Foundry, consolidating key model-development teams under CEO Eddie Wu.
By September, longtime Qwen researcher Dayiheng Liu had been appointed head of the Qwen large language model project, bringing more clarity after several rounds of reorganization.
Beyond the Model Race
Qwen had grown beyond a model family into a broader stack: foundation models for language, coding and multimodal tasks; Alibaba Cloud’s Model Studio for developers and enterprises; and consumer products such as the Qwen App.
That breadth mattered to Apple. After evaluating Chinese AI providers including Baidu, Tencent, ByteDance, Alibaba and DeepSeek, Apple chose Alibaba as a partner for Apple Intelligence in China.

By August 2026, Apple had also trained a China-specific model with Alibaba’s support, while the approved setup incorporated Qwen.
Its broader AI investment strategy also covers foundation models and AI-native applications, as well as integrating AI across existing businesses. By May 2026, Alibaba said its spending would exceed the original 380 billion yuan plan.
The Apple deal marks a new phase. The harder part may be keeping Qwen fast, focused and inventive as the machinery around it gets bigger.
DragonLab cuts through the noise to find the Chinese startups, founders, and technologies that could shape what comes next.
Have a startup we should know about? Reach us at dragonlab@technode.com.
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 TechNode(Apple’s China AI partner has had a rough ride),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。