Qwen's head Liu Dayi Heng has issued a direct denial. Future Qwen 4.0 and Qwen 5.0 series will continue to be open source, contradicting rumors that Qwen 4 would be the last open-weight line and that future models would be available only through APIs. For anyone watching the open model market, this is not a footnote: it is a strategic stance on a core asset.

The numbers help to convey the scale. The Qwen series has surpassed 3 billion cumulative global downloads, with over 460 open-source models released and more than 300,000 derivative models, holding the top spot worldwide. At this scale, Qwen models become a foundation for an ecosystem of reuse and adaptation. A move to API-only access would not just close off access to new checkpoints; it would break a chain of work already underway across hundreds of thousands of derivative projects.

The stated motivation deserves attention. Liu Dayi Heng argues that the value of AI should not belong to a few, but should become a foundational capability that users around the world can access, use, and build on. Behind that statement lies a precise industrial logic: open-weight models fuel a cycle of adoption, fine-tuning, and local distribution that a closed vendor cannot generate at the same speed. Openness here is not philanthropy; it is a lever to broaden the installed base and steer the development of the entire category.

For teams planning self-hosted deployments, the announcement reduces a continuity risk. Having access to model weights makes it possible to evaluate inference on your own hardware, keep data within organizational boundaries, and apply quantization to lower VRAM requirements. The known constraints remain: you need suitable machines, serving pipelines, and careful total cost monitoring. But weight availability is the prerequisite for even setting up a TCO comparison between cloud and on-premise. This is where openness promises have concrete effects: they change the scope of what a team can plan without depending on an external endpoint. For those evaluating on-premise deployment, AI-RADAR offers analytical frameworks at /llm-onpremise to weigh these trade-offs.

There are structural winners and losers. Organizations with data sovereignty requirements or predictable workloads can build medium-term plans on open models, without the risk of renegotiating access with every major release. Infrastructure providers and system integrators gain room, because local adoption requires optimization and management skills. API-only vendors, by contrast, must live with an alternative that keeps the exit barrier low. From AI-RADAR's perspective, this is the most interesting point: weight openness is not just a licensing question, but a factor that reshapes incentives across the entire stack, from silicon to serving platforms.

One point remains to watch: the commitment to keep models open source will need to be verified release by release, because weight availability does not automatically mean a uniform license regime. With 3 billion downloads and more than 300,000 derivative models, any change of direction would produce effects that extend far beyond Qwen, reaching an ecosystem now larger than the single vendor.