Alibaba Trials New Open-Source AI Business Model for Qwen

Alibaba is reportedly planning to implement revenue-sharing agreements for certain commercial users of its upcoming Qwen open-weight AI model. This strategic shift, detailed by sources familiar with the company’s plans, signals a move towards monetizing its advanced AI technologies beyond traditional licensing models.

Under the proposed terms, larger enterprises that leverage the Qwen model to generate revenue through their own services will be required to establish a commercial pact with Alibaba. While the precise revenue-sharing percentages are still under deliberation, the initiative aims to capture value from businesses operating at scale with Alibaba’s AI innovations.

This move diverges from Alibaba’s historical approach. Previously, while developers accessed models hosted on Alibaba Cloud under a paid structure, customers could generally deploy open-source Qwen models within their own data centers without incurring licensing fees. The new arrangement introduces a more direct commercial linkage for significant revenue-generating applications of its open-weight models.

The proposed commercial terms represent a departure from the Apache 2.0 license currently governing Qwen3’s open-weight models. The Apache 2.0 license, known for its permissiveness, allows for commercial use, modification, and redistribution, subject to its specific conditions. Alibaba’s new strategy suggests a desire for a more tailored approach to capitalize on the commercial success of its AI.

Open-source versus Open-Weight: A Nuanced Distinction

The distinction between “open-source” and “open-weight” AI models is critical to understanding Alibaba’s strategy. Open-weight models make their trained parameters accessible for download, providing a foundation for developers to build upon. However, this does not inherently guarantee unfettered access or unrestricted commercial use for all applications.

The Open Source Initiative (OSI) defines open-source AI systems as those that allow users to freely use, study, modify, and share the system for any purpose without requiring explicit permission. This definition also mandates access to crucial information, including training data, relevant code, and model parameters. While Alibaba’s Qwen models offer downloadable weights, the new commercial terms suggest a more controlled monetization strategy, especially for high-revenue applications.

Alibaba and other Chinese AI developers have been at the forefront of releasing large models with downloadable weights. In contrast, major global players like OpenAI, Anthropic, and Google primarily offer their flagship AI models through proprietary, closed systems and hosted services, emphasizing API-based access and control.

Alibaba’s proposed revenue-sharing model appears to draw inspiration from the licensing framework adopted by Chinese AI developer Moonshot for its Kimi K3 model. Released last month, Kimi K3’s license includes specific provisions for companies operating Model-as-a-Service (MaaS) businesses that exceed certain revenue thresholds.

Under Kimi K3’s existing license, businesses utilizing the model for MaaS operations must enter into a separate agreement with Moonshot once their combined revenue, along with that of their affiliates, surpasses $20 million within any consecutive 12-month period. This stipulation applies to the commercial deployment of Kimi K3 and any derivative models.

Furthermore, the Kimi K3 license mandates prominent display of the model’s name for commercial products reaching significant user bases, specifically exceeding 100 million monthly active users or generating $20 million in monthly revenue. Exemptions are provided for internal use and services offered through Moonshot’s own channels or certified inference partners.

Sources close to Moonshot’s commercial agreements indicate that these arrangements can involve revenue sharing, with some partners potentially sharing up to 30% of the revenue generated from the AI service. This indicates a growing trend among AI developers to explore sophisticated monetization strategies beyond upfront licensing.

The impact of these arrangements is already being observed. Chinese IT services firm Chinasoft International publicly disclosed a revenue-sharing agreement with Moonshot in a recent regulatory filing, underscoring the practical implementation of such models, though specific financial details remain undisclosed.

DigitalOcean Holdings, a cloud computing provider, is among the companies offering access to Kimi K3 and other Chinese AI models. CEO Paddy Srinivasan confirmed a commercial agreement with Moonshot, while maintaining discretion regarding specific terms. He characterized this approach as an “open-source freemium” model, where initial access is offered at minimal cost, with subsequent charges for larger-scale commercial use, enhanced technical support, or early access to future innovations.

The Economics of Scaling Open Models

While companies can download open-weight models without paying for API access, deploying these advanced AI systems at scale necessitates substantial investments in computing infrastructure. The operational costs associated with running large AI models are a significant factor driving new monetization strategies.

Moonshot’s Kimi K3 model, for instance, boasts 2.8 trillion total parameters with 104 billion activated parameters, utilizing a mixture-of-experts (MoE) architecture with 896 experts, of which 16 are selected per token. This complexity translates into considerable hardware demands for operators.

The demand for Kimi K3’s computational resources was so high that Moonshot temporarily paused new subscriptions in July due to strain on its available GPUs. This situation highlights the significant infrastructure challenges associated with self-hosting models of this magnitude, making cloud-based or managed services more attractive for many users.

Alibaba is employing a similar architectural approach with its Qwen3.8-Max model. This model features approximately 2.4 trillion parameters, activating around 95 billion parameters for each request, according to reports. The MoE design, by activating only a subset of experts per token, enhances scaling efficiency and manages computational load.

Cloud providers play a crucial role in this ecosystem, offering hosting and inference services for a fee. AI infrastructure companies, in turn, can generate revenue through deployment optimization and specialized services tailored to these large models.

Dan Fu, Vice President of Kernels at Together AI, emphasizes the competitive landscape where AI service providers differentiate themselves through efficiency gains in token usage and deployment optimization. He notes that “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.” This highlights the importance of practical application and value creation beyond the raw model capabilities.

The cost of developing these sophisticated models is also a major consideration. Research from Epoch AI and Stanford University indicates a substantial increase in the cost of compute-intensive training runs, roughly doubling annually since 2016. While the price of accessing models with equivalent capabilities has declined, the initial investment in research and development remains high.

When Kimi K3 was initially released, its API pricing was approximately one-third that of Anthropic’s Fable model, based on input and output token rates. However, pricing is only one facet of the total deployment cost, particularly for organizations managing high volumes of requests or operating models on dedicated infrastructure.

These operational and development costs are intertwined with the evolving licensing arrangements. Alibaba already charges for Qwen access via Alibaba Cloud. The proposed revenue-sharing model extends this monetization strategy, allowing them to capture value from companies deploying Qwen independently on their own infrastructure or through third-party services.

Moonshot’s proactive approach in attaching commercial conditions to Kimi K3, while keeping model weights downloadable, sets a precedent. The commercial agreements with DigitalOcean and Chinasoft International, though lacking public financial specifics, illustrate the growing traction of this hybrid model.

The development of these commercial frameworks occurs against a backdrop of geopolitical tensions surrounding AI technology. The U.S. has raised concerns about Moonshot’s alleged use of proprietary technology from Anthropic during its model development, a claim that Chinese officials have refuted.

The trend of releasing models with downloadable weights is not exclusive to Chinese developers. Thinking Machines Lab, a San Francisco-based AI firm founded by former OpenAI CTO Mira Murati, recently unveiled its first open-source model, signaling broader interest in this approach.

Lin Qiao, CEO and co-founder of Fireworks AI, believes there are no inherent technical barriers preventing U.S. developers from releasing more advanced open-source models. Fireworks AI collaborates with various model developers, including Moonshot, although specific commercial arrangements remain confidential.

Alibaba has yet to formally announce the definitive license for its forthcoming Qwen model or the specific revenue-sharing percentages it intends to seek from its commercial partners. The market is closely watching these developments as they shape the future landscape of AI monetization and collaboration.