Alibaba’s Groundbreaking Approach: Revolutionizing AI with Qwen Open-Source Model

Alibaba is considering implementing revenue-sharing terms for select commercial users of its upcoming Qwen open-weight AI model, as reported by Reuters based on information from insiders.

This new arrangement would entail larger companies that derive income from providing the model as a service to establish a commercial agreement with Alibaba. The specific revenue-sharing percentage is still under discussion, according to the sources.

Alibaba is set to introduce this initiative with its next open-source model. Previously, the company had charged developers for accessing models on its cloud platform, while allowing customers to deploy open-source models in their data centers without licensing fees.

The proposed setup differs from the licensing model for the current Qwen3 open-weight models, which are released under the Apache 2.0 license. This license permits commercial use, modification, and redistribution subject to its terms.

Open-source versus open-weight

Open-weight models provide access to their trained parameters for download, but this doesn’t necessarily mean that every aspect of the AI system is open or that all forms of commercial use are unrestricted.

According to the Open Source Initiative’s Open Source AI Definition, an open-source AI system should allow users to utilize, study, modify, and share it for any purpose without requiring permission. The definition also mandates access to training data, relevant code, and model parameters.

Alibaba and other Chinese AI developers have released large models with downloadable weights. In contrast, companies like OpenAI, Anthropic, and Google primarily distribute their main commercial models through closed systems and hosted services.

Alibaba’s planned terms are reminiscent of the licensing model implemented by Chinese AI developer Moonshot for Kimi K3, an open-weight model released recently. The license includes specific conditions for companies operating Model-as-a-Service businesses above certain revenue thresholds.

According to the published license for Kimi K3, a company running such a service must reach a separate agreement with Moonshot when the combined revenue of the company and its affiliates surpasses $20 million in any consecutive 12-month period. This provision applies to the commercial use of Kimi K3 and derivative models.

The license also includes a requirement for significant consumer-facing deployments. Commercial products with over 100 million monthly active users or generating $20 million in monthly revenue must prominently display the Kimi K3 name, with exceptions for internal use and services provided through Moonshot or certified inference partners.

Per sources familiar with Moonshot’s commercial agreements, these contracts may involve revenue sharing, with Moonshot potentially requiring partners to share up to 30% of the revenue earned.

Chinese IT services company Chinasoft International disclosed a revenue-sharing arrangement with Moonshot in a recent regulatory filing, without specifying the exact percentage involved.

DigitalOcean Holdings is one of the companies offering Kimi K3 and other Chinese models. CEO Paddy Srinivasan confirmed a commercial agreement with Moonshot but refrained from sharing further details.

Srinivasan described the approach as an open-source “freemium” model, where companies can access software with minimal or no initial cost before paying for larger-scale commercial use, technical services, or early access to future releases.

The cost of running open models at scale

Although companies can download an open-weight model without paying for API access, deploying large models at scale necessitates significant computing infrastructure.

For instance, Kimi K3 comprises 2.8 trillion total parameters and 104 billion activated parameters, as disclosed by Moonshot. Its mixture-of-experts architecture features 896 experts, with 16 selected for each token.

The substantial size of the model places significant hardware demands on operators. Moonshot temporarily halted new Kimi K3 subscriptions in July due to resource constraints, while Reuters reported low expectations for self-hosting a model of that magnitude due to infrastructure requirements.

Alibaba is adopting a similar architectural approach with Qwen3.8-Max, which contains around 2.4 trillion parameters but activates approximately 95 billion parameters per request, according to Reuters.

Moonshot claims that its mixture-of-experts design enhances scaling efficiency by activating only a subset of the model’s experts for each token, rather than the entire model.

Cloud providers may charge for hosting and inference services, while AI infrastructure firms can generate revenue from deployment and optimization services.

Dan Fu, Vice President of Kernels at Together AI, emphasized that companies offering AI services can differentiate themselves through more efficient token utilization and deployment optimization.

Fu highlighted the value at the application layer in leveraging models and tokens effectively to accomplish specific tasks.

The development of models poses a distinct cost challenge. Research involving Epoch AI and Stanford researchers revealed a 2.4 times increase in the cost of compute-intensive training runs annually since 2016. Stanford’s 2025 AI Index noted a sharp decline in the price of accessing models at a given capability level.

Based on published API prices, Kimi K3 was priced at approximately one-third of Anthropic’s Fable model, considering listed input and output token rates. However, pricing is just one aspect of deployment costs, especially for companies running models on dedicated infrastructure or managing high request volumes.

These costs coexist with the licensing frameworks being explored by model developers. Alibaba already charges developers accessing Qwen through Alibaba Cloud. The proposed revenue-sharing model would enable the company to generate revenue from businesses deploying Qwen independently on their infrastructure or via third-party services.

Moonshot has already imposed commercial conditions on Kimi K3 while keeping model weights available for download. Both DigitalOcean and Chinasoft International have revealed commercial agreements with Moonshot, though specifics on financial terms remain undisclosed.

These commercial agreements are evolving amid broader tensions between China and the US concerning AI technology. The White House has accused Moonshot of leveraging technology from Anthropic in developing its models, an accusation Chinese authorities have refuted.

Interest in releasing models with downloadable weights extends beyond Chinese developers. Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, launched its first open-source model recently.

Fireworks AI CEO and Co-founder Lin Qiao stated that there are no fundamental technical barriers preventing US developers from releasing more advanced open-source models. Fireworks AI collaborates with models from developers like Moonshot, although Qiao declined to discuss their commercial agreements.

Alibaba has not publicly disclosed the final license for its upcoming Qwen model or the revenue-sharing percentage it intends to request from major commercial users.

(Photo by: Alibaba)

See also: Alibaba, DeepSeek push China’s AI model race towards lower costs

Want to learn more about AI and big data from industry leaders? Explore AI & Big Data Expo events in Amsterdam, California, and London, part of TechEx alongside the Cyber Security & Cloud Expo. Click here for more details.

AI News is powered by TechForge Media. Discover upcoming enterprise technology events and webinars here.