Why Alibaba Wants a Cut of Your AI Revenue (Qwen Explained)

Chinese AI labs have spent the past two years flooding the market with open-weight models that anyone can download and run. Alibaba's Qwen family and Moonshot's Kimi series sit near the top of that pile, and both have pulled real commercial usage away from the closed systems offered by OpenAI, Anthropic, and Google. That giveaway posture is now shifting. Alibaba is preparing to attach revenue-sharing terms to its next Qwen open-source AI release, aimed at the larger companies that turn its models into paid services. The change would not affect a solo developer running the model on a laptop. It could, however, reshape how cloud providers, model-as-a-service startups, and enterprise integrators price their offerings. This post walks through what Alibaba is testing, how it mirrors Moonshot's Kimi K3 licence, and what it means for anyone deploying Chinese open-weight AI at scale.

Quick answer: Alibaba plans to require large commercial users of its next Qwen open-weight release — reportedly Qwen3.8-Max — to share a portion of the revenue they generate from it, following Reuters reporting citing two people familiar with the plans. Small developers, researchers, and internal deployments stay free. The template is Moonshot's Kimi K3 licence, which triggers a mandatory revenue-share agreement — reportedly up to 30% — once a reseller crosses $20 million in combined annual revenue.

This article covers:

• What Alibaba is changing about Qwen licensing

• How Moonshot's Kimi K3 licence became the template

• Why open-weight and open-source are not the same thing

• What the shift means for cloud providers and startups

• The bigger picture: costs, freemium, and US-China friction

• Frequently asked questions

What Alibaba Is Changing About Qwen Open-Source AI Licensing

Alibaba plans to require larger commercial users of its next Qwen open-source AI model to sign a revenue-sharing agreement. The exact rate is not yet finalized. Smaller developers and internal deployments would still be free to download and run the model. The measure is reportedly arriving alongside the open-weight release of Qwen3.8-Max, not the current Qwen3 line.

According to AI News, which relayed Reuters reporting citing two people familiar with the plans, the new terms target companies that resell the model as a paid service. Today's Qwen3 open-weight models ship under the Apache 2.0 licence, which permits commercial use, modification, and redistribution. Alibaba already charges developers who access Qwen through its own cloud platform. The new terms extend monetisation into a space it currently leaves alone: companies deploying Qwen on their own servers or through third-party providers.

The move matters because Alibaba has trained the market to expect free self-hosting. Attaching commercial conditions to that path, even at high revenue thresholds, changes the calculus for anyone building a product on top of Qwen open-source AI.

How Moonshot's Kimi K3 Licence Became the Template

Moonshot's Kimi K3, released last month, is the closest published example of the licence Alibaba appears to be copying. Companies running Model-as-a-Service businesses on Kimi K3 must sign a separate agreement with Moonshot. The trigger is $20 million in combined revenue with affiliates during any twelve-month window. Consumer products with over 100 million monthly active users or $20 million in monthly revenue face similar conditions, plus a requirement to display the Kimi K3 name prominently.

Those separate agreements can carry real revenue-share numbers. According to Reuters, reporting via AI News, one source said Moonshot can require partners to share up to 30% of the revenue involved. Chinasoft International, a Chinese IT services firm, disclosed a revenue-sharing agreement with Moonshot in a regulatory filing last month without naming the percentage. DigitalOcean confirmed a commercial agreement with Moonshot for Kimi K3 hosting, though the company has not disclosed terms.

That arrangement has been described industry-wide as an open-source "freemium" model: companies can start with the downloadable weights at little or no cost, and pay once usage or revenue justifies it, or when they want technical support and earlier access to future releases.

The Consumer-Product Clause Most Readers Miss

Kimi K3's licence also requires large consumer-facing deployments to prominently display the Kimi K3 name. Carve-outs cover purely internal use and services offered through Moonshot or its certified inference partners. That naming clause acts as free brand distribution for Moonshot at the same time it enforces the commercial ceiling.

Why Open-Weight and Open-Source Are Not the Same

Open-weight models publish their trained parameters, but that is narrower than what most engineers mean by open-source. According to the Open Source Initiative's Open Source AI Definition, an open-source AI system must let users use, study, modify, and share it for any purpose without asking permission. It must also expose information about training data, code, and parameters. Revenue-share clauses do not fit that definition.

That gap matters for how the industry talks about Qwen open-source AI. Alibaba, Moonshot, and other Chinese labs release models the wider community treats as "open." The licences attached to those releases, however, can carry commercial gates. Reading the licence, not the label, is now the practical minimum for any team building on these weights.

A useful point of contrast: Zhipu AI's GLM-5.2, released in June 2026, ships under a genuine MIT licence with no regional limits and no revenue-share clause — permitting self-hosting, fine-tuning, and commercial use outright. That makes it one of the few frontier-scale Chinese releases that actually satisfies the stricter open-source definition rather than sitting in the open-weight-with-conditions category that Qwen and Kimi occupy.

What the Shift Means for Cloud Providers and Startups

The economics look different depending on where you sit. A hobbyist or research team can still pull Qwen weights and run them on whatever hardware they can afford. A managed-inference startup routing paid API calls to Qwen open-source AI is exactly the target of the new terms.

Deployment cost is the other constraint. The table below sets out the scale involved.

Both models use a mixture-of-experts design. Kimi K3 has 896 experts, with 16 selected per token. Both can technically be self-hosted, but Moonshot itself paused new Kimi K3 subscriptions in July because usage was pressuring its available GPUs. Few customers have the hardware to serve models of that size at production scale, which is why cloud partners exist in the first place.

Dan Fu, vice president of kernels at Together AI, framed the value question this way: application-layer value comes from how a company actually turns models and tokens into something useful for its own users — not from owning the weights themselves. Companies providing AI services can still differentiate through token efficiency and deployment optimisation, even when the underlying weights come with commercial strings.

The Bigger Picture: Costs, Freemium, and US-China Friction

Training the biggest models keeps getting more expensive. According to research from Epoch AI, which analysed the training costs of 45 frontier models, the compute cost of the most demanding training runs has grown at roughly 2.4 times per year since 2016. At the same time, the Stanford AI Index 2025 found that the price of querying a model at a given capability level fell more than 280-fold between November 2022 and October 2024. Somebody has to pay for the training, and API-only pricing is not covering it as capability spreads across cheaper providers.

That helps explain why the freemium licence idea is spreading. If open-weight competitors can undercut closed API pricing while still recouping training cost from the largest resellers, the revenue-share licence becomes the obvious middle path. Alibaba is walking into that model with the next Qwen release.

Political friction hovers over all of this. In late July 2026, Michael Kratsios, director of the White House Office of Science and Technology Policy, publicly accused Moonshot AI of covertly using a technique called distillation to extract capabilities from Anthropic's Claude Fable 5 model in developing Kimi K3, and of acquiring export-restricted Nvidia GB300 hardware abroad — allegations Moonshot has not publicly addressed and which some AI researchers have questioned on technical timeline grounds.

Open-weight releases from US labs remain relatively thin by comparison, though that is starting to change. Thinking Machines Lab, the San Francisco company founded by former OpenAI chief technology officer Mira Murati, released its first open-weight model, Inkling, in July 2026. Lin Qiao, chief executive of Fireworks AI, has argued that there's no fundamental technical barrier preventing US developers from releasing more capable open-source models — the commercial and political incentives simply haven't aligned in that direction so far.

What to Watch Next

Alibaba has not published the final licence for its next Qwen open-source AI model or the revenue-share percentage it plans to seek. Two indicators are worth tracking. The first is the licence text that ships with the model. The second is the first commercial agreements the company discloses, in the same way Chinasoft's regulatory filing surfaced the Moonshot deal. If Alibaba lands its threshold near Kimi K3's $20 million line, freemium open-weight becomes the default posture for major Chinese AI releases, and US labs will face pressure to answer with something similar.

FAQs

Does the new licence mean Qwen open-source AI is no longer free?

No. Individual developers, researchers, and companies operating below whatever revenue threshold Alibaba sets will still be able to download and run the next Qwen model without paying. The revenue-share terms are aimed at larger firms that resell the model as a service.

What is the revenue-share percentage Alibaba will charge?

It has not been finalised or publicly announced. For context, the comparable Moonshot arrangement for Kimi K3 can require partners to share up to 30% of the revenue involved, according to one source cited in Reuters' original reporting.

Is the current Qwen3 line affected?

No. Alibaba has released Qwen3 under the Apache 2.0 licence, which permits commercial use, modification, and redistribution. The proposed revenue-share terms apply to the next Qwen open-source AI release, not to models already published.

Can startups still self-host Qwen at scale?

Technically yes, but models like Qwen3.8-Max activate around 95 billion parameters per request and require substantial GPU infrastructure. Moonshot's own Kimi K3 caused enough GPU pressure to pause new subscriptions in July. Most companies serving production traffic will still lean on cloud partners.

Would this count as open-source under the standard definition?

Under the Open Source Initiative's Open Source AI Definition, an open-source system must allow use, study, modification, and sharing for any purpose without seeking permission. Revenue-share requirements conflict with that. The "open-weight" label is more accurate for what Alibaba and Moonshot are shipping — unlike, for example, Zhipu AI's MIT-licensed GLM-5.2, which carries no such commercial gate.

How does this compare to OpenAI, Anthropic, and Google?

Those three primarily distribute their flagship commercial models through closed hosted APIs, not downloadable weights. Alibaba's approach keeps weights available while collecting revenue from the largest commercial users. It is a different balance point in the same market.