Alibaba Ousts Qwen's Founding Team Days After Landmark Open Source Release
The Development
On March 4, 2026, Qwen technical lead Junyang Lin, staff research scientist Binyuan Hui, and intern Kaixin Li publicly announced their departures from Alibaba via X, posting within hours of each other. The exits came less than 24 hours after the team shipped Qwen3.5—a small model series (0.8B to 9B parameters) featuring a Gated DeltaNet hybrid architecture capable of a 262,000-token context window while running on consumer hardware. The release drew praise from Elon Musk for its 'impressive intelligence density.' Contributor Chen Cheng stated publicly that 'leaving wasn't your choice,' strongly implying forced departures. Lin had grown Qwen to over 600 million downloads and was the primary liaison between Chinese AI research and Western open-source communities. Hao Zhou, a Google DeepMind Gemini team veteran, is reported to be assuming leadership. Alibaba faces its fiscal Q3 earnings report on March 5, 2026, and has recently consolidated AI efforts into a consumer-facing 'Qwen C-end Business Group' targeting wearables and hardware.
Our Take
This is the Meta Llama 4 moment for Chinese open-source AI, and enterprises running Qwen on Apache 2.0 licenses should treat it as such. Alibaba has telegraphed its intent clearly: install a metric-driven leader from a proprietary AI team, merge the model lab with a consumer hardware division, and face investors the next morning. The 'open weights as competitive moat' strategy that made Qwen a viable enterprise alternative to US proprietary models is now structurally at risk. The rumored Qwen3.5-Max moving behind a paid API isn't speculation—it's the logical outcome of every decision Alibaba made this week. For the 90,000-plus enterprises on Qwen via DingTalk or Alibaba Cloud, the licensing terms you signed for today are not guaranteed to reflect what you'll be offered when you need to upgrade.
The Exploit
Action Item
If your organization runs Qwen models in production, download and archive the current Apache 2.0-licensed weights from Hugging Face before Alibaba's March 5 earnings call resets commercial expectations and potentially triggers licensing changes on future releases.