Alibaba is reportedly preparing revenue-sharing terms for some companies that commercialise its next Qwen open-weight model. The plan would preserve downloadable weights while requiring larger model-as-a-service operators to reach a separate agreement with Alibaba.
The proposal, reported by Reuters through two unnamed sources, has not been formally announced. Neither the threshold nor the revenue percentage is final. Even so, it signals a growing licensing experiment: broad access for developers, followed by commercial obligations when another company builds a large business on top.
Downloadable weights do not settle whether a model is open source
Alibaba’s current Qwen3 open-weight models are released under Apache 2.0, which permits commercial use, modification and redistribution subject to its conditions. The reported terms for the next model would be different if qualifying businesses must negotiate permission or share revenue.
Open weight and open source are not interchangeable. Open-weight releases expose trained parameters, but may withhold training data, code or unrestricted usage rights. Under the Open Source Initiative’s Open Source AI Definition, users should be able to use, study, modify and share a system for any purpose, supported by information about data, code and parameters.
A revenue-triggered restriction may still create a useful and accessible model ecosystem, but it complicates an open-source label. The exact judgement will depend on Alibaba’s final licence, which is not yet public.
The approach resembles Moonshot AI’s Kimi K3 licence. Model-as-a-service companies whose combined revenue with affiliates exceeds $20 million over twelve consecutive months must negotiate separately. Large consumer deployments crossing specified user or revenue thresholds must also display the Kimi K3 name, with stated exemptions.
Sources cited in the article said Moonshot agreements can involve revenue sharing of up to 30 percent. Chinasoft International and DigitalOcean have disclosed commercial arrangements, but not their detailed financial terms.
Commercial scale creates costs beyond the licence
Free access to weights does not make a large model inexpensive to operate. Kimi K3 contains 2.8 trillion total parameters and activates 104 billion per token through a mixture-of-experts design. Moonshot temporarily paused new subscriptions after demand pressured its available GPUs.
Alibaba’s Qwen3.8-Max uses a similar sparse approach, with about 2.4 trillion total parameters and roughly 95 billion active for each request. Activating a subset improves computational efficiency, but providers still need substantial accelerators, memory, networking and operations at commercial traffic levels.
That creates several businesses around an open-weight model. Cloud providers sell hosting and inference; infrastructure specialists optimise deployment and token use; application companies package models into specialised services. Revenue sharing would let the model developer participate when those downstream businesses reach scale.
The economics are not one-sided. A service provider already pays for hardware, engineering, support and customer acquisition. A large revenue share could discourage adoption or push operators toward models with simpler terms. Earlier access, technical support or a differentiated model could make an agreement more attractive.
Buyers need licence clarity before building a platform
Companies evaluating the next Qwen release should wait for the final text rather than planning around anonymous-source reports. They will need definitions for model-as-a-service activity, revenue calculation, affiliates, derivative models, geographic scope and changes to future versions.
Procurement should model total cost across self-hosting, third-party inference and Alibaba Cloud. That calculation should include infrastructure, utilisation, optimisation work and any negotiated payment—not merely the absence of an API fee.
Developers also need a versioning strategy. If current Apache-licensed Qwen models and a future commercially restricted model coexist, applications should record which weights and licence govern each deployment. Fine-tunes and derived systems may introduce additional questions.
Alibaba has not confirmed the plan, so it remains a prospective business model rather than an established requirement. The company could change thresholds or abandon revenue sharing before release.
The experiment reflects a genuine tension. Model developers want broad distribution and community adoption while recovering escalating training costs. Users want the freedom and predictability associated with open software. Qwen’s final licence will show whether Alibaba can combine those goals—or whether “open” becomes a low-cost entry point followed by negotiated permission at commercial scale.
