Qwen3.5: A Leading Small AI Model, and the Team That Just Left

This is the story of Qwen. Not the press release version. The real one.

Who Built It

Lin Junyang, known online as Justin Lin, is a 1993-born PKU humanities graduate who joined Alibaba's DAMO Academy in 2019 with a master's degree. He was not a traditional AI researcher. He was a polyglot, a communicator, someone who could explain what a transformer architecture was doing in terms that a developer in Berlin or São Paulo could understand. He became the primary bridge between China's deep engineering talent and the Western open-source ecosystem.

When Alibaba first introduced the Qwen model in April 2023, Lin was the one who shaped its public identity. He gave the talks. He wrote the blog posts. He stayed up through the night during launches to answer questions from developers on GitHub and Discord. He was the reason that when a startup in San Francisco wanted to fine-tune a Chinese open-source model, they reached for Qwen instead of the alternatives.

Binyuan Hui, born in 1999, joined Alibaba DAMO Academy in 2022 after completing a master's degree at Tianjin University. He led Qwen Code, the division responsible for the model's coding capabilities. He left in January 2026 to join Meta, posting on X: "bye qwen, me too." Yu Bowen, head of post-training, departed the same day as Lin, March 4. Three of the people most responsible for what Qwen became, gone within two months.

What They Built

Qwen 3.5, released on February 16, 2026, is the final major release that Lin's team shipped before the departures. It is, by any technical measure, extraordinary.

The Qwen3.5 small model series spans four sizes: 0.8 billion, 2 billion, 4 billion, and 9 billion parameters. These are not large models. The 9B version runs natively on a standard laptop. The 0.8B version runs in a web browser. What makes them remarkable is what they can do at that size.

The models use a Gated DeltaNet hybrid architecture, combining linear attention with full attention in a 3:1 ratio. This allows a 9B-parameter model to maintain a 262,000-token context window, large enough to hold an entire codebase, while remaining efficient enough to run on consumer hardware. The models are natively multimodal, processing text, images, and video without requiring separate specialized components. They are designed for what Alibaba calls the "Agentic Inflection," the shift from AI as a chatbot to AI as an autonomous worker that can navigate user interfaces, execute code, and complete multi-step tasks without supervision.

Elon Musk, who does not typically praise Chinese AI releases, wrote on X that the models showed "impressive intelligence density." The AI Engineer community, which had been following Qwen's progress through Lin's conference talks, described the release as a blueprint for the next generation of on-device AI agents.

Alibaba has open-sourced nearly 400 Qwen models in total. The Qwen model family has become the base model of choice for a significant portion of the Chinese embodied-AI industry, for companies building robots and physical AI systems that need efficient, capable models they can run on edge hardware. Cursor, the AI coding tool used by hundreds of thousands of developers, fine-tunes and post-trains its products on top of Qwen models. The reach of what Lin's team built extends far beyond Alibaba's own products.

What Happened

The departures were not random. They were the visible surface of a deeper organizational conflict that had been building for months.

Under Lin, the Qwen team operated as what VentureBeat described as a "full-stack AI lab," a vertically integrated unit that handled everything from pre-training and infrastructure to multimodal research and coding agents. Lin had argued repeatedly, including in a public talk at the Tsinghua AI Summit in January 2026, that pre-training, post-training, infrastructure, and training teams needed to be more tightly integrated and communicate more closely. He had been building that integration from the inside.

Alibaba's leadership had a different view. The company decided to restructure Tongyi Lab, breaking the Qwen team apart into separate horizontal units for pre-training, post-training, text, and multimodal work. Lin's management scope was reduced. The vertically integrated model he had championed was being dismantled. According to people familiar with the matter, Lin's departure came as a surprise to many on the team, with one person describing a sense of regret: "It's bittersweet. He really loved Qwen."

The replacement for Yu Bowen's post-training role is Zhou Hao, a former Senior Staff Researcher from Google DeepMind's Gemini team. The appointment prompted immediate concern in the developer community. Xinyu Yang, a researcher at rival Chinese AI lab DeepSeek, posted on X: "Replace the excellent leader with a non-core person from Google Gemini, driven by DAU metrics. If you judge foundation model teams like consumer apps, don't be surprised when the innovation curve flattens."

The internal tension had a commercial dimension. Qwen's open-source strategy had made it globally influential, but influence does not directly translate to revenue. Alibaba Cloud faces aggressive competition from ByteDance's Volcano Engine, which has adopted a closed-source model strategy. In the Chinese super-app market, the Qwen app did not significantly close the gap with ByteDance's Doubao during the Lunar New Year subsidy battle. Some Alibaba executives had described Qwen 3.5, the model that drew praise from Musk and the global developer community, as a "half-finished product." The commercial objectives and the technological goals were not aligned, and when that happens in a large company, the people who built the technology for its own sake tend to be the ones who leave.

The "Gemini-fication" Problem

The pattern is not unique to Alibaba. It has played out at OpenAI, at Google, at Meta. The researchers who build something genuinely new operate with a degree of autonomy and creative latitude that is difficult to sustain as an organization grows and investors demand returns. The product managers arrive. The metrics shift from technical capability to daily active users. The people who built the thing for the love of building it find that the environment has changed in ways they cannot adapt to.

At Google, the Gemini team has been criticized for prioritizing benchmark performance over genuine capability, for building a model that looks good on paper but disappoints in extended use. The concern among the Qwen community is that the same dynamic is now coming to Alibaba, that the appointment of a DeepMind veteran to lead post-training signals a shift from research-first to metric-driven culture. VentureBeat called it "the Gemini-fication of Qwen."

For the 90,000 enterprises currently deploying Qwen through DingTalk or Alibaba Cloud, the concern is practical. They chose Qwen because it offered what VentureBeat described as "a third way": the performance of a proprietary US model with the transparency of open weights. If Alibaba decides that open-sourcing future flagship models conflicts with its commercial interests, those enterprises face a difficult choice. Industry analysts have raised the possibility that future models, including the rumored Qwen3.5-Max, may be locked behind paid proprietary APIs.

What Comes Next

Lin Junyang is unlikely to disappear from the AI world. Multiple investors and major companies had already been in contact with him before his departure, some hoping he would start a company of his own, others extending job offers. Kaixin Li, an intern who also departed, suggested that a planned Singapore-based research hub had been part of the vision: "Qwen could have had a Singapore base, all thanks to Junyang. But now that he's gone, there's no reason left to stay here."

The Qwen models themselves remain. The 400 open-source models that Lin's team shipped are still available. The 700 million downloads on Hugging Face are still there. The developers who built products on top of Qwen still have what they built. Open-source is, by design, resilient to the decisions of any single organization.

But the community that formed around Qwen, the late-night launch sessions, the GitHub discussions, the conference talks where Lin explained what the team was building and why, that community was built by specific people who cared about specific things. Wenting Zhao called Lin's departure "the end of an era." She was not being dramatic. She was being precise.

The Qwen3.5 small models are, technically, a remarkable achievement. A 9B-parameter model that runs on a laptop, maintains a 262,000-token context window, processes text and images and video natively, and is designed from the ground up to operate as an autonomous agent. The people who built it shipped it, and then they left. The models will outlast the team that made them. That is the nature of open source.

Whether what comes next will be as good is a different question entirely.