Alibaba Revolutionizes Open-Source AI with a Unique Business Model

Alibaba explores a new business model for Qwen

Alibaba, the Chinese e-commerce giant, is considering introducing an innovative business model for its upcoming open-weight artificial intelligence model, Qwen. Reports indicate that the company may implement revenue-sharing terms for certain commercial users. Two sources close to the matter revealed that this measure would primarily target large companies that generate revenue by offering the model as a service.

The idea is for these companies to enter into a commercial agreement with Alibaba. However, the specific details regarding the revenue-sharing rate remain to be defined, according to the same sources. This initiative marks a shift from Alibaba's current strategy, which allows developers to access its hosted models via its cloud platform for a fee, while permitting the deployment of its open-source models in private data centers without licensing fees.

A paradigm shift for the Qwen license

The proposed business model by Alibaba for Qwen significantly differs from the current license used for the Qwen3 open-weight models. Currently, these models are published under the Apache 2.0 license, which allows for commercial use, modifications, and redistribution under certain conditions. However, the new model may introduce additional restrictions for commercial users.

Open-source versus open-weight

Open-weight models, such as those offered by Alibaba, make their trained parameters available for download. However, this does not mean that the entire AI system is open or that all commercial uses are free from restrictions. According to the Open Source Initiative, an open-source AI system must allow use, study, modification, and sharing without requiring prior permission. This definition also includes access to training data, relevant code, and model parameters.

Alibaba, like other AI developers in China, has released large models with downloadable weights. In contrast, companies like OpenAI, Anthropic, and Google prefer to distribute their main commercial models via closed systems and hosted services. The terms being considered by Alibaba resemble the licensing model adopted by Moonshot for its Kimi K3 model, which imposes specific conditions on companies exceeding certain revenue thresholds.

The business model of Moonshot and its implications

The license for Kimi K3, published by Moonshot, stipulates that a company operating a Model-as-a-Service must enter into a separate agreement with Moonshot if its combined revenue with affiliates exceeds $20 million over a consecutive 12-month period. This condition applies to the commercial use of Kimi K3 and derived models.

Furthermore, for large-scale deployments aimed at consumers, commercial products that exceed either 100 million monthly active users or $20 million in monthly revenue must clearly display the name Kimi K3. Exemptions exist for internal use and services offered through Moonshot or certified inference partners.

According to sources close to Moonshot's commercial arrangements, these agreements may include revenue sharing. One source mentioned that Moonshot could require its partners to share up to 30% of the generated revenue.

Technical and economic challenges of open-weight models

Companies can download an open-weight model for free to access an API, but large models require substantial computing infrastructure for large-scale deployment. For example, Kimi K3 contains a total of 2.8 trillion parameters and activates 104 billion parameters, according to Moonshot. Its mixture of experts architecture includes 896 experts, with 16 selected for each token.

This size imposes significant hardware requirements on operators. Moonshot temporarily stopped accepting new subscriptions to Kimi K3 in July due to the pressure on its available GPUs. Relatively few users are able to self-host a model of this scale due to the required infrastructure.

Alibaba employs a similar approach with Qwen3.8-Max, which contains approximately 2.4 trillion parameters but activates about 95 billion parameters for each request. Moonshot claims that its mixture of experts design enhances scaling efficiency by activating only a subset of the model's experts for each token.

The role of cloud providers and infrastructure companies

Cloud providers can charge for hosting and inference, while AI infrastructure companies can generate revenue from deployment and optimization services. Dan Fu, Vice President of Core at Together AI, stated that companies providing AI services can differentiate their offerings through areas such as more efficient token usage and deployment optimization.

"At the application level, there is value in how you use it, how you actually get the models and tokens to do something useful," he said.

Financial challenges of model development

Model development presents a distinct cost challenge. Research involving Epoch AI and Stanford researchers estimated that the cost of the most computationally intensive training runs has increased by approximately 2.4 times per year since 2016. However, Stanford's AI Index 2025 revealed that the price of accessing models at a given capacity level has dropped significantly.

At the time of publication, Kimi K3 was offered at about one-third the price of Anthropic's Fable model, based on token entry and exit rates. Pricing is only part of the deployment cost, especially for companies running models on dedicated infrastructure or handling large volumes of requests.

These costs add to the licensing arrangements tested by model developers. Alibaba is already charging developers for access to Qwen via Alibaba Cloud. The proposed arrangement would also allow it to collect revenue from certain companies deploying Qwen independently on their own infrastructure or through third-party services.

Moonshot has already attached commercial conditions to Kimi K3 while keeping its model weights available for download. DigitalOcean and Chinasoft International have both disclosed commercial arrangements with Moonshot, although the financial terms have not been made public.

Geopolitical context and global implications

Commercial arrangements are developing alongside broader tensions between China and the United States regarding AI technology. The White House has accused Moonshot of using technology from Anthropic in the development of its models, an allegation that Chinese officials have denied.

The interest in publishing models with downloadable weights is not limited to Chinese developers. Thinking Machines Lab, the San Francisco AI company founded by former OpenAI CTO Mira Murati, released its first open-source model last month.

Lin Qiao, CEO and co-founder of Fireworks AI, stated that there is no fundamental technical barrier preventing American developers from publishing higher-performing open-source models. Fireworks AI works with models from developers like Moonshot, although Qiao declined to discuss its commercial arrangements.

Alibaba has not yet publicly announced the final license for its upcoming Qwen model or the percentage of revenue sharing it plans to request from large commercial users.