PLYY

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.

Alibaba is backing that push at group scale. In February 2025, it pledged at least 380 billion yuan ($53 billion) over three years specifically for AI and cloud infrastructure — more than it had spent on the two areas over the previous decade.

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.