This is a top-down adjustment.

"This is definitely a top-down adjustment by Wu Mama (Wu Yongming)." A senior source close to Alibaba told Shuzhi Qianxian about the background of Lin Junyang's departure as the technical leader of the Qwen large model. "Around the Spring Festival, Alibaba's management was engaged in a battle—organizational integration. Resources were too fragmented before and couldn't be concentrated." Just two days before Lin Junyang announced his resignation, Alibaba unified its AI brands under "Qwen."

The broader context is commercialization results. "However, technical people may not fully accept some adjustments, and they have many other opportunities," the source said. On March 5, Alibaba CEO Wu Yongming sent an internal email confirming the approval of Lin Junyang's resignation.

"Not surprising," said a senior AI industry professional upon hearing the news. "There have been whispers for months. I think this is about organizational and power adjustments. Also, there's the question of how well Qwen is performing—external and internal perspectives might differ." Another source noted that while the incident seemed sudden externally, it had been developing for months.

"It's a pity. Qwen is really good, with extensive open-source coverage across multiple sizes and diverse models. Deepseek has only one model and size, hasn't released a major version in over a year, and the open-source community mainly uses Qwen," a founder of an AI agent company told Shuzhi Qianxian. He felt Alibaba was moving too hastily, which would impact the open-source ecosystem, but acknowledged that Alibaba had open-sourced many models, unlike ByteDance, which focuses on model-based products for stronger returns. "So Alibaba needs to adjust—it's still a business after all."

"There's actually a lot of movement in the large model circle recently; people just don't know," an AI professional lamented. "There are still few truly knowledgeable people about large models in China, especially leaders."

Around Lin Junyang's departure, Shuzhi Qianxian interviewed various stakeholders to comb, sort out, organize, arrange, streamline (sort out) dynamics and competition in the AI industry. "From the current competition, 2026 might see even more drama than 2025," a senior source said.

01

Questions Alibaba's AI Must Answer Behind the Lin Junyang Turmoil

A senior observer noted that since the second half of last year, Alibaba has positioned itself against Google, "from chips, cloud, and large model capabilities, plus full traffic support." Google's stock surged after Gemini 3's launch, while Alibaba's has fluctuated.

Alibaba's management now faces two challenges: one is its e-commerce operations, still burning money in competition with Meituan in the food delivery market; the other is transforming Qwen AI into a group-wide advantage.

"While ByteDance's large model capabilities temporarily lag behind Alibaba's in some areas, its strategy and results have been better. Qwen is well-received but not translating to business success—this must be resolved," the source added.

Beyond business macro-trends, Alibaba has been restructuring its organization in recent months. In the large model era, achieving organizational "synergy" has become a core issue.

"Yao Shunyu, who joined Tencent last year to lead large model efforts, proposed at internal meetings that a company's organizational capabilities are directly reflected in its model capabilities," a source told Shuzhi Qianxian. Traditional departmental silos are exacerbated in large model competitions. Alibaba's recent organizational adjustments aim to foster collaboration, with contradictions arising during this process becoming one trigger for the resignation.

Interpretations of Qwen's performance vary. Externally, Qwen has open-sourced over 400 models with over 1 billion global downloads, making it the world's top open-source model.

However, some AI professionals observe that Qwen's open-source strategy covers all sizes—"releasing dozens at once"—with broad coverage and market share. Yet, individually, Qwen lacks a standout flagship model. "I've seen reports of Alibaba internal executives calling Qwen 3.5 a half-baked product, suggesting internal skepticism, despite it maintaining domestic leading status," an AI professional noted.

"More specifically, there's a new industry term called ARC," the AI professional continued. A stands for Agent, R for Reasoning, and C for Coding—three fierce competition arenas and lucrative directions.

Take Coding as an example: while Qwen Code upgraded to v0.5.0 by late 2025, it lags behind overseas models like Claude Code by several months. Recent model usage rankings on the overseas developer platform OpenRouter show Kimi dominating programming scenarios with a 30.8% market share, while companies like Zhipu and MiniMax are also aggressively enhancing coding capabilities.

"Critically, theoretically, once a programming model breaks through, it can achieve infinite scalability. Perhaps Alibaba judges that future directions could lead to unified models, eliminating the need for numerous 'messy' small models," a senior agent professional added.

This incident also ties closely to Alibaba's reflection on its open-source strategy. Several sources told Shuzhi Qianxian that in 2023, Baidu led the domestic market with a closed-source strategy, while Alibaba surpassed it in 2024 by pursuing both open-source and closed-source approaches. In 2025, Deepseek's emergence strengthened open-source advocacy. However, after a year, Deepseek failed to release major updates, while closed-source Doubao achieved commercial success, showcasing strong market competitiveness; Zhipu and MiniMax also went public.

This has sparked industry reflection: a balance between open-source and commercial strategies is needed.

"Do we really need to develop so many small models? Should we focus more on deep R&D for core models? The current market scenario seems unfavorable, requiring all parties to re-examine development strategies," an AI professional said, emphasizing the need to reassess whether technological innovation investments match outputs.

Lin Junyang initially led open-source efforts at Qwen and was rapidly promoted from P6 to P10 over four years after key members like Zhou Chang left. The new hire from Google has a reinforcement learning background, similar to Yao Shunyu, who joined Tencent last year—both aligning with AI's current agent explosion phase.

An external pressure driving this adjustment is competition with ByteDance. Comparing their rivalry around the Spring Festival, sources note ByteDance now has Doubao App at the frontend, models like Seedance providing tools, traffic support, and token-based closed loop (closed loops).

A senior cloud computing professional added that ByteDance took 1.5 to 2 years to grow DAU for public cloud calls to nearly 50%. While calls currently contribute minimally to cloud revenue (single-digit percentages), the trend suggests exponential growth, akin to electricity usage. "If Alibaba loses this market, it's unacceptable," explaining Alibaba's goal to capture 80% of new market growth.

02

Alibaba's Critical Battleground in 2026

From Alibaba's management perspective, a senior source judged the current core task is making AI genuinely serve Alibaba's broader strategy. As e-commerce extends into local life services, expanding traffic scale, leveraging AI (including agents) to build competitive advantages at traffic entry points becomes crucial. "After all, Alibaba leads competitors in this area."

However, he noted that Qwen's attempt to subsidize milk tea orders requires further observation of its long-term effects and retention. During the Spring Festival, major companies adopted App-era traffic subsidy tactics to compete for large model users, "but I believe models fundamentally provide capability value."

"Competition among giants reflects how last-generation traffic and this-generation model capabilities form chemical reactions," he said. Finding naturally aligned partnerships remains a key exploration for giants.

From the AI market competition, the senior source observed several major developments in the past 1-2 months: first, coding agents have advanced to senior programmer levels, applicable in serious business scenarios. Second, Claude CoWork (the industry's term for general-purpose agents) can generalize capabilities by stacking skills, ultimately covering and replacing most white-collar jobs. Additionally, Openclaw (Xiaolongxia) gained popularity, drawing widespread attention.

"These three are closely linked: CoWork represents the integration of general-purpose agents and skill models, while Xiaolongxia validates CoWork's value in specific applications, allowing personal assistants to expand functions infinitely and solve problems autonomously. By 2026, the 'lobster model' will evolve continuously, even expanding into enterprise markets due to its high alignment with human habits—its Github stars already exceed Linux."

The three hottest sectors in 2026 will be: AIGC, especially video entertainment content; coding; and Xiaolongxia.

How does this connect to Alibaba? This week, Alibaba Cloud launched Agent QoderWork, rivaling Claude CoWork, and offered first-purchase discounts for Coding Plan. Alibaba's subsequent moves in "ARC" and industry competition trends warrant further observation.

Beyond this, AI cloud is rapidly rising in Alibaba's cloud business arena.

"The current landscape feels like 2013 again, when public clouds first emerged as a battleground," a senior third-party cloud computing professional lamented, predicting a significant increase in users choosing public clouds.

Previously, private cloud costs weren't high—deploying 10-20 machines might cost 1-2 million yuan, whereas now a single machine exceeds 2 million yuan, "easily" leading to billion-yuan investments. "Not something ordinary enterprises can afford." He predicted that if the previous split was 60% public cloud and 40% private cloud, in the AI cloud era, it might shift to 10% private and 90% public.

In this context, full-stack optimization for public clouds and standardizing AI cloud products become competition focal points.

In the first round of AI cloud competition, ByteDance used its Doubao and traffic apps to drive initial AI cloud token calls, gaining an edge in the public cloud token ecosystem. Alibaba currently lacks sufficient C-end traffic, "so this adjustment means Alibaba must proactively focus on product-side efforts." Meanwhile, enterprise markets also offer opportunities, such as Claude's cloud-based enterprise service model.

Regarding the open-source issues triggered by Alibaba's incident, an AI professional predicted that Chinese companies, including Alibaba, will continue advancing open-source strategies to compete with Europe and America (European and American) developed markets for users, especially amid current overseas application surges. Several sources told Shuzhi Qianxian that in the large model arena, open-source vs. closed-source is no longer critical—model capability matters most. "As long as it's useful, people will call it."

The 2026 competition around resource integration, business innovation, and user experience will intensify, making the market even more lively.