For anyone tracking Chinese AI labs, Qwen was a reliable name in open source—until recently. The departure of Junyang Lin, the lab's prominent leader, coincided with a sudden silence: the entire 3.7 model line remains fully closed, with no public weights, no repositories, and none of the transparency that marked earlier releases.

The departure of Junyang Lin and the Qwen 3.7 vacuum

The firing made headlines, but the real story is what followed: nothing. No checkpoints, no accessible model cards, no option for an organization to download and run the model on its own servers. Qwen 3.7 exists, but only behind APIs. For a lab that built its reputation partly on weight availability, this marks a strategic pivot—one many observers read as a shift toward monetizing inference rather than enabling an ecosystem.

China’s open-source map: everyone but one

A glance at recent release dates makes the anomaly clear. GLM shipped version 5.2 in mid-June, Kimi open-sourced K2.7-Code days before, MiniMax followed with M3, Step released 3.7-Flash in late May, MiMo published V2.5-Pro in April, and DeepSeek dropped the V4-Pro/V4-Flash duo on April 24. All open, all available for self-hosting. As of today, Qwen is the only major Chinese lab absent from this list—the last to keep its latest line behind closed doors.

What on-premise deployments lose

For organizations operating in regulated environments, with data residency requirements, or simply with a zero-cloud strategy, the absence of an open-source LLM from a major player is far from a footnote. Closed models force API calls to external data centers, multiply privacy risks, and block any form of fine-tuning on proprietary data. Moreover, the Total Cost of Ownership of an API-based approach can quickly become unsustainable as inference volumes scale. Having access to a model like Qwen 3.7 locally—quantized and optimized for consumer GPUs or on-premise servers—would have been a concrete alternative. Its absence nudges users toward solutions that, however performant, give up direct control over the infrastructure.

A market demanding transparency

Qwen’s decision highlights a growing tension between labs’ commercial logic and the requirements of mature enterprise adoption. The paradox is that even as global regulations push for auditability and digital sovereignty, some of the most advanced players are withdrawing direct model access. This is not an isolated dynamic; it’s a signal that AI governance also runs through code availability. For anyone evaluating their stack today, the lesson is clear: reliance on a single vendor, especially without public open-source commitments, is a risk to be calibrated carefully. China’s ecosystem remains rich with alternatives, but Qwen’s disappearance from the open-source table is a warning that no deployment strategy should ignore.