Alibaba's Qwen revenue share: open weights now have terms

Alibaba plans to make large commercial users of its next open-weight Qwen model hand back a share of the revenue they earn from it, Reuters reported on August 7, with the terms landing alongside the open-weight release of Qwen3.8-Max. The rate is not final. The direction is the story: open-weight AI licensing is turning into a commercial term sheet, and a portability plan built on the word "open" is thinner than it looked last quarter.

What actually happened

Alibaba's previous posture was straightforward. You paid when you ran Qwen on Alibaba Cloud. If you pulled the weights and ran them in your own data center, you did not pay — Qwen3 shipped under Apache 2.0, which permits commercial use, modification, and redistribution. The proposed change targets companies making money by offering Qwen as a service, and Reuters says negotiations over the percentage are still open.

There is a template already in the market. Moonshot's license for Kimi K3 requires anyone selling K3 as a service above roughly $20 million in annual revenue to strike a commercial agreement, and Reuters reports that agreement can run to a revenue share of up to 30%. Alibaba appears to be following the same shape.

Two caveats worth stating plainly. This is a Reuters report citing people familiar with the plan, not a published license — until the release lands with a LICENSE file in it, the specifics can move. And the thresholds under discussion are aimed at model-hosting businesses, not at the company running an internal support agent on forty seats.

Why open-weight AI licensing matters for your business

Almost nobody reading this crosses a $20 million model-resale threshold. That is not the risk.

The risk is that "we can always self-host an open model" has been the standard escape hatch in every vendor conversation for two years, ours included. It is why we tell clients to keep the model layer swappable. That argument quietly depended on open weights meaning MIT or Apache — no royalty, no revenue test, no counterparty. If the category drifts toward tiered licenses with revenue triggers, the escape hatch still exists but it now has a contract attached, and contracts change on the licensor's schedule.

The second-order effect reaches you faster than the first. Your AI vendors — the inference host, the agent platform, the coding tool with a "bring your own open model" tier — are exactly the businesses these terms target. A 30% revenue share on a hosted open model does not stay on the host's income statement. It shows up in your per-token price at renewal.

So do the unglamorous thing this month. Write down every model you depend on, and next to each one put the actual license name and the date you last read it. Not "open source" — the SPDX identifier. Then check whether your inference provider is reselling under a term you have never seen. Portability was never about the weights being downloadable. It is about knowing, before you need to move, exactly what you are permitted to do and what it will cost.

Sources: Reuters, via Quartz, AI News.