The timing was almost surgical in its cruelty. On the morning after Alibaba's Qwen team shipped Qwen3.5's four compact models, 0.8B to 9B parameters, capable of vision understanding and reasoning mode-switching, runnable on a consumer laptop with 7GB of RAM, outperforming models several times their size — the project's technical lead posted a five-word farewell on X: "me stepping down. bye my beloved qwen."
Junyang Lin was not leaving by choice. His colleague Chen Cheng confirmed it within hours: "leaving wasn't your choice." She had been up with him the night before, pushing the Qwen3.5 release live.
It was clear Lin was not alone. Binyuan Hui, another core lead — the kind of engineer still coordinating model deployments at 6 a.m. Beijing time — was also gone. Shortly after, Kaixin Li, a key contributor across Qwen3.5, Qwen VL, and Qwen Coder, posted his own sign-off: "Signing off from Alibaba Qwen. Grateful for the chance... Onwards and upwards." At least three to four core contributors departed in a single day. The entire technical spine of one of the world's most respected open-weight AI projects had been quietly severed.
The Corporate Logic Behind the Wreckage
What happened inside Alibaba Cloud is not officially confirmed — the company has issued no statement — but the picture assembled from multiple firsthand accounts is damning in its coherence.
Alibaba Cloud had begun evaluating the Qwen team using daily active user metrics: the currency of consumer apps, not foundational AI research. A team that had built one of the most downloaded model families on Hugging Face, beloved by academics, developers, and open-source communities worldwide, was being benchmarked against ByteDance's Doubao — a domestic consumer chatbot. By that measure, Qwen looked like a failure.
The restructuring that followed was textbook corporate politics. A replacement lead, reportedly sourced from a non-core role on Google's Gemini team, was installed. New management began reporting directly from the team, bypassing Lin — not firing him outright, but rendering him structureless and irrelevant. You Jiacheng stated plainly: "Alibaba Cloud kicked out Qwen's tech lead." Several insiders converged on a single word for what transpired: politics.
The replacement's background, reportedly in reinforcement learning, struck observers as misaligned with Qwen's multimodal, open-weight research agenda. Community skepticism was immediate and vocal.
What Is Actually Being Lost
Qwen was not just an AI product. It was a philosophical commitment — a rare instance of a frontier lab delivering open-weight models competitive with the world's best proprietary systems, then releasing them publicly. Martin Ziqiao Ma described Qwen as the bridge between frontier labs and the external research community. That bridge, he said, may now be broken.
The fear articulated by Yuchen Jin, CTO of Hyperbolic Labs, is pointed: Qwen may cease delivering frontier open-source models entirely, pivoting instead toward closed, commercially monetized architectures. If that happens, a model family that powered an entire ecosystem of fine-tuning, research, and local inference disappears from the commons.
The community's retort spread fast: "Qwen is nothing without its people."
Jiaxi Cui's verdict was blunt — Qwen3 may stand as the team's final masterwork for the foreseeable future. After this, the roadmap turns commercial.
The Permanent Tension
This is not a novel story. It is, in fact, the foundational tragedy of open-source AI inside large corporations: the moment a research project achieves enough influence to matter, it becomes subject to the metrics of a business that never fully understood what it had built. Downloads, citations, and developer goodwill do not appear on Alibaba Cloud's KPI dashboards. DAU does.
Where Lin, Hui, Li, and their colleagues land next is unknown. Speculation ranges from founding a new independent lab to joining open-source-aligned organizations. Whatever comes next, they carry with them the demonstrated capacity to build something remarkable from scratch.
We thank them for one of the greatest contributions not only to open-weight AI, but to the whole of humanity. We do hope they will continue working on cutting-edge open-weight LLMs in a new way — because the world needs people who build for the commons (open-weight LLMs), not just for the quarterly report and moat.