Two Qwen3.8 models shipped in August 2026 and both download freely. One is Apache 2.0. The other carries a license that can require you to negotiate with Qwen before any commercial use.
The flagship Qwen3.8-2.4T-A95B declares its license as qwen3.8-max in the model card metadata rather than a standard identifier. The smaller Qwen3.8-27B declares apache-2.0. Same family, same month, two different sets of obligations.
That matters because the previous generation set a different expectation. Qwen3.5-35B-A3B, from February 2026, is plain apache-2.0.
One number will not keep you compliant
The license file contains two conditions, and they do not share a unit, a window, or a consequence.
| Clause | Trigger | Window | What you must do |
|---|---|---|---|
| Attribution | 100M monthly active users or US$20M monthly revenue | Monthly | Display the model name prominently in your product |
| Separate license | US$50M revenue, licensee plus affiliates | Any consecutive 12 months | Obtain a license from Qwen before commercial use |
So compliance is not one number to watch. It is a monthly measure for attribution and a rolling twelve-month aggregate for the gate.
The gate is a negotiation, and it counts your parent company
It aggregates affiliates. The text reads "the licensee or any of its affiliates" and totals revenue across them, so a small subsidiary inside a larger group can sit above the line from the first commit.
It is scoped by business type, not by volume. The trigger is conducting "a Model as a Service or AI Work Assistant business", so a company selling model access is treated differently from one using the same weights internally at the same revenue.
It is a gate, not a royalty. The requirement is to "obtain a separate license" before commercial use. The text names no percentage and no rate, which means the number you would pay is not knowable from the license itself. That is a different kind of risk from a published revenue share: it is a dependency on a negotiation that has not happened yet.
One detail matters if you audit dependencies by metadata: the license file in this open-weights repository is headed "Qwen3.8-Max License", borrowing the name of Qwen's proprietary tier. A scanner keying on repository names or license identifiers will not tell you which of those two things you are agreeing to. Open the file. Dependency metadata has been wrong before on this site: an MCP setup guide pointed at a package that was never published, and the registry was the only thing that said so.
Only two kinds of team need to care
All of the above comes from license text and model-card metadata read on August 20, 2026, with nothing downloaded, run, or benchmarked. Both threshold clauses were confirmed still present in the published LICENSE file on the day this article went out, by a check that runs before the site builds. So this says nothing about which model is better, and it is not legal advice: "affiliates", "Model as a Service", and "AI Work Assistant business" are defined inside the license, and if you are anywhere near a threshold that text and a lawyer are your sources.
What the reading does settle is who needs to care. If you are self-hosting for internal use and nowhere near the thresholds, both models are usable and the flagship's conditions never activate. If you are building a product that resells model access, the aggregate revenue of your whole corporate group is the number to check before you fine-tune, not after.
And if your requirement is that no future revenue milestone can change your terms, the 27B is the only one of the two that gives you that, because Apache 2.0 has no thresholds to cross.