Alibaba has unveiled its most powerful AI model to date. Qwen3.8-Max has 2.4 trillion parameters. This places it just below the Kimi K3, introduced last month by industry peer Moonshot AI. According to Alibaba, the model immediately ranks among the world’s best on multiple AI leaderboards.
According to Reuters, with this launch, Alibaba is further intensifying the competition among Chinese AI developers. Companies such as Alibaba, Moonshot AI, and DeepSeek are rapidly trying to outdo one another with ever-larger open-weight models. In this race, not only performance but also efficiency is of great importance. The goal is to keep powerful AI accessible, which ensures that the required computing power and operational costs do not become unmanageable.
Top rankings on AI lists
Qwen3.8-Max has 2.4 trillion parameters, the values that an AI model learns during training and uses to recognize patterns and perform tasks. This keeps the model below the 2.8 trillion parameters of Moonshot’s Kimi K3. However, that number alone says little about the model’s ultimate performance.
According to Alibaba, the new model immediately scores highly on Arena.AI, a platform where AI models are compared by the community. Among text models, Qwen3.8-Max is the best-performing Chinese model. Globally, it is surpassed only by the latest models from Anthropic. In the field of image analysis, it ranks second, also behind an Anthropic model.
Like Kimi K3, Qwen3.8-Max can process text, images, and video. It also supports a context window of up to one million tokens. This allows entire software projects, extensive legal files, or hundreds of pages of documentation to be analyzed in a single pass.
Mixture-of-experts approach aims to reduce costs
Despite its enormous size, the entire model need not be active for every task. Alibaba uses a so-called “mixture-of-experts” architecture. This approach deploys only the specialized components required for a specific task. According to the company, approximately 95 billion parameters are activated per request. This is expected to significantly reduce both response time and costs.
Alibaba also claims that Qwen3.8-Max independently completed a software engineering project in sixteen days. Details regarding the scope or complexity of that project have not been disclosed.
The model will be made available next week via Alibaba Cloud Model Studio. This will allow developers to deploy the system for their own applications. The launch underscores once again that Chinese AI companies aim to further strengthen their position in the global open-weight AI market. Meanwhile, U.S. companies such as OpenAI, Anthropic, and Google do not disclose the size of their closed models.