Chinese AI companies are fundamentally changing how open-weight models work by attaching commercial strings to freely available code. Rather than simply releasing model weights for anyone to use, Alibaba and Moonshot AI are now requiring large companies that profit from these models to negotiate revenue-sharing deals, a shift that could reshape the entire open-source AI landscape.

What Are Open-Weight Models and Why Do They Matter?

Open-weight models make their trained parameters available for download, allowing developers to run them on their own computers or servers without relying on a company's cloud service. This contrasts sharply with how OpenAI, Anthropic, and Google distribute their main commercial models, which are accessible only through closed, hosted services.

The appeal is obvious: developers get powerful AI without vendor lock-in or ongoing API fees. But the definition of "open" is becoming contested. Under the Open Source Initiative's Open Source AI Definition, a truly open system should allow users to use, study, modify, and share it for any purpose without seeking permission. However, open-weight does not automatically mean every part of the AI system is open or that all commercial uses are unrestricted.

How Are Chinese Companies Monetizing Open-Weight Models?

Moonshot AI's Kimi K3 model, released last month, pioneered this hybrid approach. Under its published license, companies operating model-as-a-service businesses must reach a separate commercial agreement with Moonshot once their combined revenue exceeds $20 million during any consecutive 12-month period. The license also requires large consumer-facing deployments exceeding either 100 million monthly active users or $20 million in monthly revenue to prominently display the Kimi K3 name.

According to sources familiar with Moonshot's arrangements, those agreements can include revenue sharing of up to 30 percent. Chinasoft International, a Chinese IT services company, disclosed a revenue-sharing agreement with Moonshot in a regulatory filing, though it did not reveal the percentage involved. DigitalOcean Holdings, a major cloud infrastructure provider, also confirmed it has a commercial agreement with Moonshot but declined to provide specifics.

Alibaba plans to introduce similar revenue-sharing terms with its next Qwen open-weight model, though the exact percentage has not been finalized. The arrangement would differ from the current Qwen3 open-weight models, which are released under the Apache 2.0 license permitting commercial use without licensing fees.

DigitalOcean Chief Executive Paddy Srinivasan described this approach as an open-source "freemium" model, where companies can access software at little or no initial cost before paying for larger-scale commercial use, technical services, or earlier access to future releases.

Why Are These Models So Computationally Demanding?

Both Kimi K3 and Qwen3.8-Max use a mixture-of-experts architecture, which activates only a subset of the model's parameters for each request rather than running the entire system. Kimi K3 contains 2.8 trillion total parameters but activates only 104 billion for each query, according to Moonshot. Qwen3.8-Max contains about 2.4 trillion parameters but activates around 95 billion per request.

This design improves efficiency, but the models still require substantial computing infrastructure when deployed at scale. Moonshot temporarily stopped accepting new Kimi K3 subscriptions in July after saying usage had strained its available graphics processing units (GPUs). Reuters reported that relatively few users are expected to self-host a model of that scale because of the infrastructure required.

How Are Companies Creating Value Beyond Just Hosting?

Cloud providers and AI infrastructure companies are finding multiple revenue streams beyond simply charging for model access. These include:

• Hosting and Inference: Cloud providers charge for running the model and processing user queries through their servers.
• Deployment Optimization: AI infrastructure companies generate revenue from optimizing how models run on different hardware and reducing computational waste.
• Efficient Token Use: Companies can differentiate by using fewer computational tokens to deliver the same results, reducing costs for end users.

"At the application layer, there's value out there for how you use it, how you actually get the models and the tokens to do something useful," said Dan Fu, vice president of kernels at Together AI.

Dan Fu, Vice President of Kernels at Together AI

The pricing landscape reflects these dynamics. At published API prices at the time of release, Kimi K3 was priced at about one-third of Anthropic's Fable model based on listed input and output token rates. However, pricing is only one part of the deployment cost, particularly for companies running models on dedicated infrastructure or handling high volumes of requests.

What Does This Mean for the US-China AI Competition?

These commercial arrangements are developing alongside broader tensions between China and the United States over AI technology. The White House has accused Moonshot of using technology taken from Anthropic while developing its models, an allegation Chinese officials have rejected. Treasury Secretary Scott Bessent warned in July that Chinese firms could face financial sanctions and Entity List blacklisting over what he characterized as model theft.

Interestingly, ByteDance, the company behind TikTok, stands apart from these accusations. ByteDance founder Zhang Yiming reportedly banned researchers from distilling rival AI models in 2023, years before US sanction threats made such abstention politically useful. Distillation involves feeding queries to a powerful model, collecting the outputs, and using them to train a cheaper system that approximates the original.

ByteDance is conspicuously absent from Anthropic's accusations. The US lab alleged that DeepSeek, Moonshot AI, and MiniMax used roughly 24,000 fraudulent accounts to generate more than 16 million exchanges with Claude in a coordinated extraction campaign. Alibaba's Qwen operation was accused of the largest known effort, some 28.8 million exchanges using nearly 25,000 fraudulent accounts.

The timing of ByteDance's 2023 policy raises questions about whether the company demonstrated unusual foresight or genuine research principles. Beijing is pressing its labs to build better models with less compute, a demand shaped by export controls rather than preference. Distillation is an obvious shortcut under those conditions, which is what makes refusing it costly.

Despite its caution, ByteDance's best models have been surpassed by smaller Chinese labs. Its Doubao-Seed-Code model and Trae coding tool have underperformed against Zhipu's GLM 5 and Moonshot's K2. However, the commercial picture is less bleak than the benchmarks suggest. Doubao reached 345 million monthly active users by March, well ahead of Qwen at 166 million and DeepSeek at 127 million, and ranked second globally by app users behind ChatGPT.

The strategic question now is whether Chinese labs can sustain the pace of model improvement without borrowing from American systems. ByteDance has effectively volunteered to answer it through its abstention from distillation. Whether that choice proves to be principle, foresight, or an expensive mistake depends on results nobody can yet measure.