Alibaba Cloud opens access to a multi-model builder for developers

Alibaba Cloud (阿里云) has made a move that could significantly reshape the landscape of AI coding tools in China. The company has updated its subscription-based Coding Plan service, adding support for multiple competing models: its own Qwen 3.5-Plus (千问3.5), as well as GLM-4.7 from Zhipu AI (智谱) and Kimi-K2.5 from Moonshot AI (月之暗面). Now developers with a single subscription can freely switch between models, choosing the optimal tool for each specific task.

The launch of yet another AI service subscription may not surprise anyone — the market is oversaturated with offerings. But here, the architecture of the solution matters. Alibaba Cloud is not simply selling access to its own model; it is building an aggregator platform where models from direct competitors sit on the same shelf as the proprietary Qwen. This is fundamentally different from the approach chosen by most vendors: typically, each company promotes exclusively its own model, locking users into their own ecosystem.

Technically, Coding Plan integrates with key AI coding tools: Qwen Code, Claude Code, Cline, and OpenClaw. This means a developer can work in a familiar environment, while the subscription acts as a unified backend, providing access to different models through a standardized interface. The inclusion of Claude Code in the list of supported clients is noteworthy — it shows that Alibaba is not attempting to isolate users from Western tools; rather, it aims to become a universal inference provider.

The choice of models is also deliberate. Qwen 3.5-Plus is the latest iteration of Alibaba's flagship model, which in recent benchmarks demonstrates strong results in code generation and analysis. GLM-4.7 from Beijing-based Zhipu AI (a company closely tied to Tsinghua University) is renowned for deep contextual understanding and work with long code fragments. Kimi-K2.5 from Moonshot AI, a startup, is a model that has gained enormous popularity in China for its ability to process ultra-long contexts, which is critical when working with large codebases. Together, these three models cover different use cases, and the ability to switch between them without additional costs creates real value for professional developers.

This solution reflects a broader strategy. Alibaba Cloud, Asia's largest cloud provider, is clearly making a bet not on the model level, but on the infrastructure level. The logic is simple: models are becoming increasingly interchangeable, leadership in benchmarks shifts from one player to another every few months, but the infrastructure through which these models are consumed cements long-term relationships with customers. This is the same logic Amazon Web Services applies with Bedrock, and Microsoft with Azure AI Studio, offering access to models from different vendors through a single platform.

For the Chinese AI coding market, this is an important signal. Until now, competition among Alibaba, Baidu, ByteDance, Zhipu AI, and Moonshot AI has been conducted primarily at the model level — each company sought to prove that its solution generates code better than the others. Alibaba Cloud's move shifts the focus: the company is effectively saying that the struggle for the best model is less important than control over the distribution channel. If developers become accustomed to obtaining all models through Coding Plan, Alibaba Cloud will become the intermediary through which a significant portion of AI coding in China flows — regardless of which model currently leads.

The question of pricing and access terms remains — Alibaba Cloud has not yet revealed details of the updated tariffs. But the very fact of aggregating competing models under a single subscription sets a new standard for the market. Other cloud providers in China — Huawei Cloud, Tencent Cloud, Baidu AI Cloud — will either need to replicate this move or convincingly explain why tying to a single model is better than the freedom to choose. In an era when no single model maintains leadership for more than a few months, the arguments in favor of a multi-model approach look increasingly compelling.